posted 03-02-99 11:39 PM ET            

How could the AI in SMAC perform more poorly than in it's predecessor engine? I think I know. Sid and Brian's additional layering in this game outpaced their refinement of the AI. And I also see that AI refinement as pretty limited. I'll lay out some concrete analysis.

The AI has to do essentially four things: manage bases, manage units, manage the society and interact with opponent factions.

Base management: We see the AI in action with the auto governors. Notoriously awful. We also see it in action when we penetrate data links. It's laughable at times. Very little needs to be said on this point. The infrastructure management is so poor that even the generous industry advantage the AI enjoys at transcend level can't make up for the inefficiency of their use. Special note here: I turn off the random events now, just to give the AI half a chance. The AI seems to give no priority to building improvements that prevent bad random events. I watched factions get repeatedly ravaged by viral outbreaks, etc. in the first two games before deciding that it was really giving me an unfair advantage.

Unit management: We all know about the wandering do-nothing auto formers. That's the enemy former AI in action. And in SMAC, as in CivII, the formers mindlessly lay down roads in every tile, while profitable improvements go undone. Colony pods wander for ages without planting bases. Bse location has taken a step backward. Then there's the military strategies. Same old CivII predisposition for the ground assault, the pilliage stop that leaves units in open ground, the air/missile attacks that come *after* the conventional unit attack instead of before, etc. I've seen the AI build drop units, but I've never seen a drop attack. I've seen the AI "hold" bomber units in cities when they could have (and should have) been used for counter-assault. The litany goes on and on, but the most important point is that you'll never lose a war if you remember two things: (i) the AI is an imcompetent defender, and (ii) the AI won't attack your bases as long as your attacking it's. The rest is just execution.

Society Management: As far as I can tell, this task has been greatly oversimplified in the name of faction fiction, i.e. each faction has a pet social management scheme and tends to stick to it. This may contribute to game atmosphere, but any human player who optimizes social engineering even only every now and then is getting a big leg up on the machine.

Diplomatic Management: This is one area where I think we were promised a lot, but got little. Even more to the point, the improved relations I have with the AI factions are tremendously key to a very effective divide and conquer strategy. Unlike in CivII, alliances with the AI actually work. Unfortunately, that turned out to be an imbalancing feature of the game.

Unit Design: This is a special case. The auto unit designer, which seldom designs useful units, is (again) the AI in action. At first, I loved the unit design feature, but it's also the biggest game imbalancer in more ways than one. First, drone control on upper levels is a snap, once the Police power can be added to any unit (not that the AI ever seems to build Police units). Second, the AI can be suckered into building expensive and useless unit designs. Example: hit some AI bases with needlejets and rovers until you encounter your first AAA and/or ECM units. Stop producing air/rover units. Start producing and stockpiling native life form units and infantry. Wait a few turns and unlease your new units against a host of expensively rush produced or upgraded AI units designed for yesterday's war. Sure, the AI "smartly" reacts to your attack patterns, but that can be abused to even greater advantage.

posted 03-03-99 12:01 AM ET            

Micromanaging terrain: Terrain management time has been exponentially increased over CivII by the redesign of terrain features and the addition of several layers of improvement options. I spend a *lot* of game time on land management (a factor contibuted to not a little bit by the color scheme). Given the extremely limited turn times under which I expect this game to be managed in 'net encounters, I'm wondering how I'll ever manage to get anything done on this basis alone.

Micromanaging Units: The unit design workshop is going to be a real curse in managing live games. Every time I discover a new conquer tech, I wind up spending 5-10 minutes in the workshop. That's going to be unacceptable live. Use the auto designer? That's just plain unacceptable!

I can't see! I can't hear!: So help me, I *love* the window dressing in this game, but how can it do anything but get in the way of multiplay? It's impossible to tell one unit from another without clicking on them. It's going to be especially bad trying to figure out which units are which in the context of a multiplayer game.

Finally, getting away from the time issues to the game play issues. I look at this game and see a design especially susceptible to the "rush" strategy that turns virtually all RTS's into glorified Doom fests. At any given point in the game, right on up to the top of the tech tree, attack numbers are double defense numbers, and there are always ways around base defenses available to the attacker (which is why any marginally competent human can smoke the AI in war as long as he stays on the attack). That spells "rush".

posted 03-03-99 09:27 AM ET            

Regarding the AI, I think Firaxis may have given it their best try. I'm not sure it is possible to give an AI intelligence that could surprise you, or actually fight off your attacks as described above. I'm not sure I could.

posted 03-03-99 12:02 PM ET            

Analyst: As much as I hate to admit it your points regarding the AI very accurate and well stated. I wish to add just a few UI issues to the list. Before I do so let me preface my comments with the flame-proofing statement that I do love the game and have yet to be able to put it down (metaphorically of course).

1. I am bothered by the former-oriented nature of terraforming. Someone, I forget who, brought up a far superior solution. They suggested an option to define the actions you would like to have performed on a given terrain square and then have available formers drawn to that square to do the necessary terraforming actions. I am a programmer so I have SOME idea of how hard this could be, but I think it MIGHT be possible. If it is possible to fix but wasn't done(which Firaxis would be far more aware of than I) I would say it tops my list of problems with the UI.

2. The failure of some supposedly available features, like the bombing run, also bothers me. I'm confident this will be fixed in a future patch so I am not terribly concerned.

3. Finally I'll take my life in my own hands and admit that the graphics do kind of bug me. The unit graphical differences appear to only be based on armor/weapon/chassis combinations. A few special abilities come with minor adds to the graphics yet some of them which should look radically different (submarines anyone)look so similiar as to be meaningless. As much as I feel the unit graphics could be improved I would rate this fix as minor at best. I would LIKE it if this was improved, but I'm perfectly capable of shift-Rclicking the square and reading the unit descriptions.

All that mentioned I will say that I enjoy this game much more than CIV II. Even the flaws in the AI do not detract from the fact that I prefer it to the 'bad AI-cheat for higher difficulties' strategy of CIV II. All in all "Bravo Sid & Brian. I love this game dearly. And I hate/love you both for taking my life away.....again!"

posted 03-03-99 12:42 PM ET         

That's the only one I've seen. It certainly doesn't use them to the extent that I do, but in this extreme situation, it did do one.

posted 03-03-99 02:00 PM ET            

The problem with game AI is always going to be a problem. The top computer scientists at MIT, Carnegie Mellon, Stanford, among other institutions, spend 40 years and countless grant dollars on AI research, yet they haven't come up with anything intelligent enough to qualify it as artificial "intelligence". Now you expect some guy can just whip together a computer game that would constantly beat the infinitely adaptable, always learning human? On a PC? (no offense, Brian. ) Perhaps you'd cite Deep Blue beating chess world champion as a counter example, however it mainly relies on brute force search. That's why it sits on the massively parallel IBM box costing multi-million dollars. For a game the complexity of SMAC, that strategy wouldn't work. The best one can do, other than some simple path-finding algorithms, is use heuristics or rules to guide the AI. The problem is, no matter how many rules you put in, humans are always going to figure it out sooner or later. In the end, the "unbeatable AI" is going to be an unachieval goal in any forseeable future (well, unless we had a breakthrough and discovered the Secrets of Human Brain ).

posted 03-03-99 02:50 PM ET            

Single player: The AI is, at times, grossly inefficient, just as you said. Even with their growth and industry bonuses, they STILL can't keep up. Inadequate auto-formers and wandering pods are probably the biggest culprits, along with the incompetint Auto-govs.

Multi-player: Here's where I disagree. I think that the AI is the weakest link in SMAC's chain. Put a real person that knows what they're doing behind the wheel, and games will likely be very interesting. As for rushing, this will undoubtably be the case in the early stages of any MP game, but ONLY if two players are started close to each other. Geographical seperation should prevent any early rushing, in general (it just takes too many resources to build up a transport fleet big enough to overcome an adequate defense)
You're right about micromanagement, though. Each turn just takes too long, especially in latter stages. PBEM is a solution, but games could last for months.

Aside from all this, though, the game's great
(a little buggy, though)

posted 03-03-99 03:48 PM ET            

This is a very persuasive criticism of SMAC. I have noticed obvious inefficiencies in the CP management of resources. It's possible that some of these questions could be addressed by the design team, although as a programmer I know that it can be extremely difficult to get a computer to act intelligently under the variety of circumstances possible in a game of SMAC.

Let me just say, though, that you are obviously in the top percentile in terms of skill. I have very challenging games at the Librarian level, and based upon the forum posts, there seem to be many people who find the game challenging.

I'm curious as to whether you tried playing with each faction? For me, University of Planet fits my playing style (massive tech plus base building) to a tee, so playing as the Believers makes it MUCH more difficult.

posted 03-03-99 04:02 PM ET            

1. Is the SMAC AI dependent on CPU speed or anything like that? For example, does P2 450 run with better AI than a P133? If so does this impact your situation Analyst? (I'm reminded of Total Annihilation where I wasn't synced to lesser equipped opponents and blew them away with nukes while they were still building infrastructure...)

2. Not being the most up to date on AI development, is it not more likely that game companies have trouble creating "dumber" AI? Simply put, SMAC is a math problem right? Firaxis or whichever game company creates AI by placing algorithms akin to human decision making within the confines of the game system. I'm sure the problem is not making smarter AI (who would merely crunch numbers and kill you likety-split) but dumber, more human-like AI.

posted 03-03-99 05:11 PM ET         

1) Yes, base govenors and auto formers are pretty retarded. No real advance here over Civ. I never used them then and I've never used them now. In fact, I've never considered them an "AI" issue (in my view, AI has always referred to military/political/economic/diplomatic gameplay), but now that you mention it, this should be the simplest kind of AI to implement - especially base management. No improvement here over Civ, and I wasn't surprised.

2) Hence micromanagement follows from the defects of 1) above. This not only slows down gameplay - which is clearly slower than in Civ - but does indeed make multiplayer hard to imagine. You'd have to require that all players turn on base governors and autoforming in order to speed up play, with the ironic result that we'd all be as dumb in these areas as any AI player that might be in such a game (heh!).

3) On the diplomatic AI, here I would differ, especially on whether it unbalances the game toward the player vis-a-vis Civ. In Civ, once you reached a certain critical mass, the diplomacy would basically come to an end, kinda like a bad marriage. From then on, you knew it would become either a city-by-city steamroller crunch, or a nuke the suckers' capital before their spaceship arrives kind of ending.
Although increased hostility is still the general trend, the AI now has more leeway because there are more ways to fight and win the game in SMAC, while at the same time this is the easiest thing to implement since the overall decision tree is simple compared to military actions or terraforming. That's why Civ's retardedness in this area was so inexcusable and why this is one of the most improved features in SMAC, albiet still slanted towards the inevitably ever-growing hostility of the AI (I mean at transcend level) in order to make up for its equally inevitable deficiencies in the other areas.

4) There have been improvements in military AI, but it still lacks STRATEGIC military sense. I've yet to see it conduct multi-tiered offensives on a broad front. One example: the Hive predictably carries out the "serial city roll" with a gang of units on a narrow front. They also have the habit of "homing" that army in the newly conquered city. Knowing this, if I have the money and am in proximity, I bribe (mindcontrol) the city (all the easier becaus e it hasn't been converted and is probably in drone riot to boot) and get a big fat Hive army in the bargin. This was a favorite tactic in Civ2 as well.
Tactically, I've yet to see: a) cooridinated naval task forces; b) cooridinated amphibious assaults; c) coordinated air strikes to soften up a target for ground assault. I HAVE seen: a) coordinated air strikes; b) correct defensive build response to previous assault; c) cooridinated shock troop/artillery assaults; d) non-suicidial assault behavior (I once held off a large Gaian assault force with 3 artillery batteries + defensive units when they realized the futility of ever getting near enough to my base in sufficient strength to carry the assault - after raking thier encircling lines a few times, they retreated).
All of the negatives and none of the positives were in Civ.

posted 03-03-99 05:37 PM ET            

Analyst: Excellent points about the A.I, most of which I find to be correct. I haven't played much Transcend yet, but I'm disappointed to learn how poorly it performs. As for librarian level, well, it's just as well I enjoy empire building as well as military conquest.

Some other points that need to be reiterated:

* The unit combat animation is pretty disappointing. One animation for every weapon type - even CivII had more variety in the sound department. A suggested fix would be to have at least four different animations (+ psi), for instance machine gun, laser, missile/explosive, and plasma, and use a firepower system to further change the strength of each.

* The point and click interface wastes a lot of time. Having a means to 'view' the square under the mouse simply by moving the pointer over it would make terraforming and unit selection a lot simpler. And having a proper stack dialogue window (ala Warlords), and the ability to designate certain units and cities with a 'goto' number, RTS style would also save a lot of time. I estimate it would be around 30%.

* Some balancing issues seem to be unresolved (as usual, this early after a game's release). The Believer's tech penalty makes them much weaker than other factions (tried winning transcend with this faction, yet, Analyst?). The production cost of a needle jet is almost identical to that of an equivilently armed rover, despite the major difference in value between them. And, as analyst pointed out before, the game does lend itself to a rush mentality (particularly with the cheap needle jets. Once I get these, the game is virtually over.)

posted 03-03-99 07:29 PM ET            

1) The field of Artificial Intelligence sprung up in the mid-50's. It hit it's highpoint in the late 60's and hasn't achieved that much since then. In the 60's and early 70's, brilliant scientists at top institutions came up with all kinds of different human modeling algorithms for simulating human computing. NONE of these have ever succeeded. They've tried all kinds of techniques such as Production Rules, Agents, and Constraint Satisfaction, among others. None have worked.

2) One of the biggest problems with AI, as any philosophy major will tell you, is that we don't have a definition for it. A person can make a solid arguement debating whether even HUMANS are intelligent, much less a computer. Some people see intelligence as based on representation(this symbol means this, this symbol means that); Some others believe that intelligence is grounded in knowledge and facts; etc. etc. How can you make a computer think if you don't know how you think? A famous scientist by the name of Turing developed what has become the standard intelligence yes/no test...Place a person in a room with two terminals. On the end of one terminal is a human and at the end of the other is a "thinking" computer, neither of which the original person can see. The original person then types questions into their terminal to see if they can determine which one is human and which one is a computer. Except for one very specialized case, this test has NEVER been passed.

3) Lastly, the Big Blue chess player. In the 50's and 60's, researchers thought that a computer would overtake the worlds chess champ within 5 years or so. They thought that was easy; having virtual supercomputers in millions of homes across the globe talking to each instantanious? That was unheard of! When IBM finally beat Kasparov(right? I forgot his name), it wasn't by building an intelligent computer. Big Blue could look, I don't know, millions of moves ahead, while Kasparov only looked about 20. Kasparov crushed Big Blue the first game and was playing a much better game the second time than the computer was. Then he made a mistake. This cost him the game and because he was so frustrated, he also lost the next. The computer can't get frustrated and so wasn't affected. So that makes the computer more intelligent right? I don't think so. Some of man's greatest achievements are because of mistakes. Anyone want to teach a computer to make mistakes? BTW, Big Blue is still much lower ranked than Kasparov. Most chess analysts/players would agree that had they kept playing Kasparov would have soundly defeated Big Blue.

Anyway, the point of this is, think about all the options in SMAC. I've yet to hear anyone complain about a lack of options in SMAC. Sure, a little thing here or there maybe, but no one has denied that there are thousands of options in this game. Trying to get a simple piece of silicon to simulate the most impressive computer the world has ever or will know for quite some time now, the brain, is very improbable. The best we can hope for is that Sid and Brian keep tweaking and releasing patches. It sucks for us but it's the only viable method.

posted 03-03-99 07:44 PM ET            

An interesting thought on Turing Test would be this. Let you play a multiplayer game with a bunch of human and computer opponents, but no chatting is allowed. If you can't tell which factions are controlled by human, which by computer, then SMAC would have passed a partial Turing Test.

posted 03-03-99 09:00 PM ET            

On the subject of improved A.I
One notable cheat to improve the A.I that Blizzard used for Starcraft was to allow the A.I to view the human player's build queues, as well as having a completely revealed map to start with (the fog of war is left on, however).

This has two advantages, in that it helps the A.I to plan ahead and find better locations to place bases, and it's invisible to the human players, unlike the enhanced production facilities.

posted 03-04-99 09:22 AM ET            

Agoraphobe: From you're post, I think that there's something you don't understand about how SMAC (and any other similar game) works. When you automate a function, what happens is that the game AI takes over that function. That's the exact same game AI that runs the enemy factions. Consequently, when tons of people complain that the auto governors are idiots and the auto formers are wasteful and grossly inefficient, whether they understand it or not, they're also commenting on the quality of enemy AI terraforming and base infrastructure building as well. That's what I mean when I say that observing the "auto" functions is observing the enemy AI in action.

To all of those who think that my comments are, in some form, illegitimate because I'm an above average player: Any improvement in the game helps everyone at every level, not just me. One day, those of you who are challenged at Librarian level will be at the right hand side of the learning curve too. My contention is that you'll be there quicker then you realize (and quicker than you necessarily should) due to the AI weaknesses. A better AI is better for all of us.

In general, I would agree with those of you who are observing that the AI shows visible improvement in the conduct of single base tactical engagements. Other improvements are also evident, such as better pre-attack massing, keeping it's military technologically up to date and base targeting. In my games, however, the AI seldom has an opportunity to show off these improvements because I keep it on the defensive. I have observed zero improvements in that area, save for the "reactive" behavior I described earlier (and since that can be turned decisively against the AI by shifting strategies, it's a step backward, not forward). The AI still has no concept of feinting maneuvers or counterattacks. Indeed, it often wastefully holds/fortifies offensive units when put under the gun, which is one of the many reasons to keep it under the gun. For those of you who wonder how I can be as good as I claim, a big part of the answer is: offense, offense, offense. There's a strong correlation on these boards between people who are impressed by the enemy AI and people who say that they like to pursue building and growth and are indifferent to conquest. Sure, if you let the AI build up a huge empire and a huge army, it can show off some new gimmicks, but if you want to win, don't ever let it get that capability.

But all of that focuses on conduct of war, which really misses the boat, it's conduct of peace where the AI loses to the human big time. I'm not a computer programmer, but I find it hard to believe that it requires breakthrough advances in artificial intelligence programming to "teach" this program to terraform tiles sensibly, build useful base improvements and do other things of that sort. I think there's probably a much simpler explanation here than that it's an impossible achievement of programming. I think that there was a development emphasis on adding new layers and dimensions to the game that was very successful. I think there was a development emphasis on creating a compelling scenario and scifi backstory for the game that was also very successful. I think there was a concerted effort to improve the AI's conduct of war tactics that had limited success. And I think that somewhere, in all of these priorities, there wasn't enough focus on: (i) the AI's ability to lay the brick and mortar foundation of it's empire and (ii) a human beings ability to use the newly developed features of the game (especially the diplomacy and unit design features) to abuse the AI in really nasty ways.

The result of all of this for me is, as I said, a game that's easier to beat at the highest level than it's predecessor. I just can't believe that was Sid and Brian's intended result, so I'm trying to give some feedback on why I find that to be the case, in the humble hopes that this feedback will contribute to a better SMACII. My relationship with Sid's games goes all the way back to playing Pirates! on a Commodore 64. This is the first time I've ever felt like the difficulty of the AI has taken a step backward.

To answer the question posed by Brother Greg (and several others), I've never found another TBS gaming experience that holds a candle to Sid's games. In terms of strategy games, it's either his stuff or go back to playing chess against the computer as far as I'm concerned.

posted 03-04-99 12:42 PM ET            

My first and most important observation is that it does not posess the "look ahead" move algorithim that, say, Deep Blue has. This is evident in that it cannot "anticipate" an assault like a human player can.

Second, programming the AI for even a chess program is an enormous programming challenge.
I don't need to prove that SMAC is much more complicated than the simple 2-dimensional game of chess (for humor's sake, imagine a game of chess on a 500 x 250 board, complete with bodies of water, transports, bunkers, and technologies (drop singularity pawns, anyone?) ) Since the best programmers in the world cannot even develop a chess program that can beat a human chess champ, it's asking too much to ask little old Firaxis to program Deep Blue 2 into their AI algorithim for a game which is vastly more complicated than chess.

I think there is definately room for improvement, but don't expect anything substantial (else Firaxis could quit making games and make a TON of money writing AI algorithims)

You mentioned that you try to keep the computer on the defensive; I'm a pacifist SMACer, so the AI does get the opportunity to build up an army and launch an assault. I HAVE observed some very well coordinated attacks, but when I launched a counter-attack, the AI couldn't handle the pressure. I noticed that it is geared to respond well to a constant stream of attacks (noting its ability to customize defensive units to your attackers), but it simply cannot anticipate the quick, crushing attacks that scoop up a dozen cities in two turns.

posted 03-04-99 01:13 PM ET            

To amplify a point that was just mentioned, I've noticed that for as well as the Hive can handle individual attacks on its cities, if I can manage to take, say, 3 cities in one turn the Hive AI seems to disintegrate. It will fly dozens of planes around my units without actually attacking anything. It will move lots of offensive units into the open and then just have them stand there for multiple turns while my aircraft pick them off.

This led me to thinking about how I would try to program and debug the AI for SMAC. Assuming that units are assigned goals (defend this city; attack that city; etc), there should be a cheat mode which allows the human to select an AI unit and see what its current goal is.

If Firaxis doesn't already have this mode, it would certainly help them. But it would also help us help Firaxis, because when the AI does something we don't understand ("Why the hell are those units standing around the city waiting to be picked off?"), we could turn on the cheat mode and see what the units are up to. This would enable us to provide much more substantive feedback to the AI development.

We would be able to say, for example, "When a single faction has taken over one or more cities on a Hive-owned island, the Hive AI is still concentrating on defending a city that is not under attack. Instead, it should treat this as an emergency and reassign all offensive units to take back the critical cities."

posted 03-04-99 02:20 PM ET            

We now find ourselves in a position to question the Firaxis people in two main areas in regard to the obvious and numerous problems with A/I. BTW, although I am not a programmer, I do have a background in machine automation that required computers to visually identify aircraft parts and perform numerous operations such as drilling, reaming, deburring and riveting. Therefore, I think I am well qualified to discuss this issue even though I can barely beat my computer chess game on the easiest level.

First, the obvious problems with things like the autoformers should be very easy to fix. The computer should be able to assess the terrain and assign point values to the terrain building potential of each square. They should then be able to prioritize building funtions after cross referencing with the population assignments on what is produced. Therefore, if the base is producing 0 nutrients, the A/I should then direct an autoformer to build a farm to allow the nutrient number to become positive. Yeah, Yeah, Yeah, I know this will be much harder to implement than to hypothesize. However, if you sell games to make a living, don't be surprised if people point out your weaknesses and are disappointed by your results.

This same argument applies to the many other mechanics problems and I think similar solutions could be found for mindless unit movements and even the governors.

Second, we have the larger problem with the overall model or game engine that is used. Some of you have compared this to using a chess algorithm and I actually think that this is probably not the way to go. The reasons are simple. In chess, you have a defined playing field with defined moves and very limited abilities. As such, this is referred to in engineering terms as a "static" condition. As such, inputs and results are highly predictable and this allows for planning the downfall of your opponents. Therefore, the person who posted the allegory of planning 7 moves ahead of a six move opponent is excellent. BTW, It is also why I get drubbed in chess games.

However, TBS games do not use this model. Rather, a TBS environment is "dynamic" as the situation is constantly changing. I will not go into specific points to prove this as all of you who have preceeded me have done a good job as to why.

Therefore, you have 2 choices. You can either try to build a static model to approximate the dynamic one, or put in the extra effort to build the dynamic model. It is my hunch that the Firaxis crew are still using the static model. As such, you will always have an A/I that is reactive and one that will never be able to "predict" the moves of a human player. In short, it is a huge advantage to the human player and the person who sees this should almost never lose. This can be said for even a new player at the highest difficulty.

I can prove this very easily. You attack with aircraft and the A/I builds AAA troops. You then build worms or psi attack units and roll the AAA infantry. This is additionally enhanced by the ability to modify units in the field. Also, my points are proven by the numerous "CHEATS" that are built into the code. For example, I would expect that the A/I knows he's getting rolled by airpower not by your force mix or the units that he is in contact with, but rather by what you are building. It would be interesting if Brian would comment on this. Also, you can see the numerous cheats by the lack of randomness in the Wrath of God events and how the various factions respond to you when you are too far ahead or behind. I won,t go into all the other cheats as you know what they are and how they work.

As a result, we have what we have. Although I will concede that the A/I is better than any other game to date, it also results in a very predictable and disappointing game for many of us. Therefore, I think that it is fair for us to Challenge Firaxis to build a better game or for them to come flat out and say that this is all the have.

posted 03-04-99 02:25 PM ET            

Deadron, that was a very insightful observation about the AI imploding when you deprive it of multiple bases in a short time. I'd never put my finger on it before, but you are absolutely right. Your observations also reminded me of one of my own: as far as I can tell, as long as you own a single base on an AI "home" continent/island, you are immune from an assault against your own. That is to say, that the AI approach to warfare is single minded. If you have a land presence on "their" land, then the AI obligingly restricts it's concept of the "front" to that presence. A corralary of this rule is that, as far as I've observed, the AI never carries out an amphibious assault as long as there exists any land bridge connecting your land holdings to theirs.

posted 03-04-99 03:46 PM ET         

My educated guess of order of complexity:
Diplomacy (simpliest)
Base-building
Terraforming
Military (most complex)

That's why one should _expect_ improvements in diplomacy as a matter of course (and demand more!), while applauding _any_ improvement in the military action.

Hence the base-building logic is probably the biggest hole in the game AI, if my assumptions on complexity are correct. Thus any improvements in this area would produce a relatively bigger payoff in game play for work involved in implementing them. You'd achieve a) a stronger computer AI player, who'd last for the long-haul on a firmer build foundation, as Analyst noted, and 2) faster gameplay for the human, who would trust all of the micromanagement to the governor. If there ever is a SMAC v3, it's on build and terraforming AI that it should focus its efforts.

posted 03-04-99 05:44 PM ET            

Even Deeper Blue has these problems; while it has been taught a lot of chess theory, and can make simple positional choices when they are obvious, it lacks Gary's inestimable ability to invent the kind of unforseeable game breaking moves he is so good at.

So while the computer can look much further ahead than Gary (don't kid yourself, the machine will wipe the floor with ANY human in this department), it isn't looking for the same thing as the human player. Deeper Blue is searching for the best available material strategy, and consequently it will pass over a move that Gary might have played, in favour of one with a more obvious (material) payoff. And lose as a result.

How this applies to SMAC is important, because the rules are fundamentally different. SMAC is inherently a material game, at least as far as the Terraforming/base management aspects go. I also use a simple formula when terraforming - food first (make sure a base has at least +4 surplus), then minerals, then energy. Modified by the location of the former at that time - i.e moving 3 squares to make a mine is less efficient than moving on to plant a solar array. Paying attention to the material value of each square, it is fairly easy to come up with an algorithm for terraforming a base.

Analyst's rush strategy, however, is a positional matter, and might well be impossible to solve (by the way, I hate playing like that, it ruins the personality of the game. But it's very effective, I must admit. The Believers sound like your perfect faction, in fact, analyst ). Knowing which city to attack, with what unit combination, how to place that combination, and when to do it is something that a machine will struggle with to no end, while an experienced human will be able to sort through logically. Again, however, the A.I can be taught simple strategies, and respond with a material focus that will be enough to defeat maybe 90% of players. However, as has been shown, these tactics can be taken advantage of, as well.

posted 03-04-99 05:54 PM ET            

The primary difference between humans and computers in chess, and the primary AI problem that has never been solved, is that the elite human chess players will only consider a handful of "candidate" moves in any given position, but may analyze (or attempt to analyze anyway) their impact in move sequences 20 moves deep or even more. [Amateurs also only consider a handful of moves, but usually not the same ones as better players, contrary to what you suggest, and certainly never analyzing their consequences as deeply as better players.] Computers have proven incapable of engaging in that "candidate" move step of chess analysis, so chess AI has only improved to the extent that processing power provides the ability to spin out raw calculations of all move possibilities. It has been calculated, however, that there are 10 to the 14th power potential variations on just the first ten moves in a chess game. Even Deep Blue can't handle that kind of calculating job. That's why it's programmers limited it to 6 moves of depth under actual game time limits.

This actually has relevance to SMAC in that in SMAC, as in chess, what seem like intuitively obvious "moves" to us may be impossible to describe to the AI in terms of code. The more "candidate move" possibilities you present the AI, due to the complexity of directions offered by multiple layers of features, the more liklihood that the AI will choose a poor set of "candidate moves" to calculate and proceed from. Consequently, I accept that much of what seems simple to us may not be so simple to code as AI. My suggestion would be pretty simple--if adding multiple new features paralyzes AI decisionmaking ability, it might be best to leave some features out of the game (the unit design workshop would be one such candidate in my estimation). It would be my contention that if you're goal is to create a great single-player gaming experience (which is my understanding of the primary goal of the makers of SMAC) then tensions between more features v. better AI performance ought to be resolved in favor of better AI performance.

posted 03-04-99 06:54 PM ET            

First things first. To clarify a few names, the computer that beat chess world champion Garry Kasparov is called "Deep Blue", you can go to http://www.chess.ibm.com/ to read about it.

In my opinion, comparing chess to SMAC is not entirely approriate. The game of "Go" is probably a closer analogy. In case you don't know, Go is an ancient oriental game where you have a 19x19 board, two people each puts black or white stones on the intersection. If your stones got encircled, take them out. In the end the one with bigger territory wins. While the Big Blue has come up with the computer and program that can actually beat chess world champion, no one has attempted the same thing for Go simply because of the complexity of the game. I think SMAC is on the same level of complexity, if not higher.

Fortunately, unlike in Go where you have to have a grand strategy and everything is correlated, in SMAC we can apply the CS101 trick: divide and conqure. The former, for example, is not strongly related to other aspects of empire building, so we can study its algorithms separately.

Man, all this theory talk gets my head spinning. Without further ado, I present you my Gaian forest-road-sensor auto-former psudo code:

1. Look at neighbor squares. If found hostile units (native mindworms, enemy units), move to the opposite direction. End Move.

2. Look at current square. If within my city radius and forest is plantable (not rocky, not xenofungus), plant forest. End Move.

3. Look at neighbor squares. If one is within my city radius and forest is plantable, move there. End Move.

4. If current square is in-between my two cities (i.e. the combined distance from this square to the two cities is less than or equal to the direct distance between those two cities), build road. End Move.

5. If current square is not in-between my two cities (the opposite of 4) and sensor can be build, build sensor. End Move.

6. Look at neighbor squares. If there is one that satisfies 4, move there. End Move.

7. Look at neighbor squares. If there is one that satisfies 5, move there. End Move.

posted 03-04-99 11:16 PM ET            

One thing I noticed is that the AI does not build up an invasion force until it has declaired vendetta upon someone. Perhapse they could feint peace until their invasion force is just off your coast, and THEN declair war (just as I have often done). In short, the AI is not as underhanded as it should be when it comes to diplomacy.

posted 03-04-99 11:19 PM ET            

And, hey, I don't *have* to be a warmonger. I just finished a game as Morgan a few minutes ago. Transcend, Iron Man, Standard sized Random map. Diplomatic Victory. Elected Supreme Leader in 2305. I'll say this: I should praise Sid and/or Brian (or whoever came up with the idea of Diplomatic Victory) with great praise; praise them with great praise. It saves me hours and hours of hunting down the enemy to the last unit and base to finish a game. That was a really tedious element of Civ II--spending a dozen extra hours winning a game I knew I'd won. Diplomatic Victory is a great way to relieve players of hours of tedium.

This was a game that demonstrated the unbalancing power of alliances. I had a pathetic starting location, being cheek to jowl with UofP in a corner of a landmass. I could only build four land bases before I ran out of space. So, naturally, I steamrolled them and wiped them out, using an alliance with UofP's opposite neighbor, the Hive. That brought me face to face with the Hive. Fortunately, I'd come into possetion of mindworm breeding technology very early in this game and was able to breed worms to use against Yang (a very effective way around those perimiter defenses). Yang, who had wiped out Lal, himself, surrendered surprisingly early in this fight. That gave me a pact brother who would focus on building military.

posted 03-05-99 07:53 AM ET            

All we need for a challenging adaptive AI is for it to use the tit-for-tat strategy that's so effective in elemntary game theory.

Problem is, how does the AI in SMAC recognise the what the human player has done so successfully in order to imitate it? That is, how does it separate the essential from the extraneous? What model does it use to distinguish the two?

Regarding Yang, his stats do seem too strong: -1 energy per square will slow him at the start but it is quickly overcome by his increasing industrial might, particularly in human hands.

Player1 hasn't seen the AI build up troop concentrations on the border before declaring Vendetta, but I've suffered that on multiple occasions, so it can be as sneaky as player1 wants.

Another tactic it uses is to gradually increase the number of border incursions, apparently to see how much I'm willing to tolerate. If I advance my own troops to impede the incursions, then the AI stays peaceful, but if it gets away with it without response from me, then it judges me a weakling and attacks in force.

The remarked-upon concentration on home defence by an AI that faces enemy attacks on its own soil is strange, because the MOO2 AI had no such limitation. A typical AI tactic in MOO2 was to concentrate a fleet at a forward base, just waiting for an opportune moment. This often came in the form of my moving a fleet to attack one of the AI's systems. At the exact turn when I was too far away to turn back in time, the AI would launch its attack, guaranteeing that I would lose a valuable planet.

Perhaps one of the problems with SMAC AI is that it has no concept of a fleet? (Or army group.) But I thought that was one of the promised features?

Experienced players easily beat the MOO2 AI's tactical engine, but the fact remains that it has one. Moreover, it's one that knows some ways in which to combine weapons effectively: for example, it knows that in order to board it has to first immobilise your ships.

It's a great pity that the Simtex team which has been formally dissolved don't offer their services as either employees or contractors to a gameplay-oriented firm such as Firaxis. Their experience would be a valuable complement.

As for SMAC's extra features being a handicap to the AI, that's our fault. We the keen players and netizens wanted these mountains of wonderful customisations, and Firaxis obliged.

Maybe what we needed were more Analysts, before the game was designed.

Personally, I most like a game that is elegant, where the complexities are derivative of a small set of simple rules. (Not quite as elegant as Go, tho.

The Design Workshop is an idea in that direction, and for humans it works - modulo the annoyance of having to flick it on and off about three times more often than is logically needed.

Which raises the question of modularity and its relation to analysis. At the obvious risk of boring CS 101 students to tears, an easy to use system is one that consists of well-made modules, each with a simple but adequate interface to the other modules with which it must communicate directly. A complicated, irregular, counter-intuitive interface which changes state when you least want it to, is the antithesis of that.

Unfortunately, we see examples of such interfaces in several of the windows in SMAC. For instance, the Current-Build, the Build-Queue, and the Design-Workshop don't work together seamlessly as they should; instead, the City-Window pops up between them, when we don't want it.

A modular approach to Exploring an area, Building a base, or Mounting an Attack or a Defence, would improve the game's algorithms significantly. For example, building a force according to the principles of combined arms - as mentioned, a concentrated air attack followed by a heavy ground assault - would break many a human city.

Especially if the same tactic - and that's all it needs to be, one known effective tactic - were applied to multiple cities in the same turn. Then the human would find the shoe on the other foot. After all, this isn't a secret tactic invented yesterday, nor is it one that's difficult to implement - otherwise I surely wouldn't be able to do it!

The obvious assumption in a simultaneous attack is that each air/navy/army combined assault team has a well-defined goal of taking city X.

We also need a higher level of military group, to satisy the higher goal of taking as many cities from the same player this turn as possible.

As military analysts will attest, timing the arrival of each team at their various goal simultaneously is not always an easy task, but even if the timing is out by a turn or two it can often be difficult to fend off an approximately coordinated assault, as a defender at a similar technological level has comparable logistic constraints.

As resources permit, other military coordination groups (groups of teams that are separated in position but united by a plan, albeit a simple plan) may be on different missions in the same turn. This would often enable a resource-strong AI player to overrun two or more opponents in the same turn.

Incidentally, it was stated that one can afford to let an AI player grow unchecked. This isn't always so in MOO2, so I don't see why it need be so in SMAC, especially if the AI is improved to use military teams right.

posted 03-05-99 12:15 PM ET            

Zoetrope, you made some fine points in that post. True, the various modules in SMAC may not be extrememly efficient in some areas. But to restructure them would almost certainly require a major overhaul, and I'm sure that's not what Firaxis wants right now.

posted 03-05-99 01:08 PM ET            

I almost gave up the game in despair right that moment. Fortunately my HQ had finished building Aerospace Complex and was turing out elite copters like there was no tomorrow. Within a few turns my copters swept the invasion force, even took back the lost city. Still, the believers' invasion had such an impression on me I dreamed about men in orange suit chasing me.

Morale of the story? This excellent show of AI tactics must have been hard coded or scripted, for in comparison Mirian made no use of Air Power I had given her. Otherwise she could have easily countered my copters.

Unless Firaxis can come up with a consistent AI model, we are going to see situations like this. AI performs briantly under certain circumstances, simply because the strategy is programmed in, yet overall it comes up short against human opponents.

posted 03-05-99 01:16 PM ET            

Pragmatist, last I checked, there was only one Supreme Leader :-) I understand your point, but when I called the election the Gaian's were fielding chaos/plasma units against my tachyon/probability units. Plus, my conquered pact-brother Yang was actually fielding Planet's largest military, also as technologically up to date as my own. Deirdre wouldn't have had a prayer. The Diplomatic Victory option just saved me hours of tedious slaughter. The game was not in doubt. I didn't need to do any more to feel I'd earned the full flush of victory. Again, I praise with great praise whomever got Diplomatic Victory into this game, saving me hours of tedium in endgame mopups.

Zoetrope, I have no idea what you just said but go you programmers go! Posts like this can't help but stimulate thought at Firaxis about how to improve a notably weak AI and clunky interface design.

posted 03-05-99 05:43 PM ET         

Question to Analyst: did you do much boreholing, echelon mirroring, condensor building or fungus-destroying in that game?

Hmmm, are the factions just plain-vanilla cutouts at bottom?

In the next (imagined) release of SMAC or whatever, may there be an "Analyst" population type (along with Workers, Drones, Empaths, etc.) that gives your base/city +3 AI for base governers and autoforming, yuk, yuk...another plug for my big playability payoff / relative ease of implementation request for more intellegent base/terraforming AI.

posted 03-05-99 09:48 PM ET            

Perhaps a second 'Social engineering' type menu could be added. Call it 'Planet enginneering, and allow you to specify the priorites for your cities and formers. For example:

Growth - city govs build Crechs, Hab complexs, Rec commons, police units - anything to accomdate a large population. Formers build farms w. sensors or solar collectors and at least one condensor per base. The governor trys to have at least two formers in action at any time, and will also be more willing to 'engineer' golden ages.

Energy - city govs build energy banks, Fusion labs, etc, and turn surplus citizens into engineers. Formers build farms and solar collectors, or forest where this is more efficient (i.e low altitude barren squares). At least one borehole is constructed.

Labs - city govs build Network nodes, fusion labs, and, may start secret projects devoted to science, such as the Supercollider. Surplus citizens are converted to Librarians or Empath/Thinkers later on. At least one borehole is constructed.

Industry - governor builds factories, skunk works, and facilities to minimise ecological damage, and improve forest output, while formers build forest, mines, and boreholes wherever possible.

Military (defensive) - the governor focuses on defense, and will build perimeter defenses, crechs, lots of garrisons, and facilities to enhance military units. Any formers will operate under industry rules, maximising the mineral output of the city, but with new emphasis on building sensors and clearing fungus.

Military (assault) - the governor will build high quality assault units, using the relevant city structures. Formers operate as for military defensive.

BTW, when selecting a goal, you would get a message asking whether you want the formers for that city to come under governor control.
This way, the formers AND the city are working cooperatively, and hence former behaviour is governed by the needs of the city.

Each of these can obviously be modified further for individual cities, but at least having these categories allows the A.I to set a goal for the city/former combination. For instance, you could adopt a 'ruthless' approach to an occupied city, allowing your governor to build punishment spheres, or a 'liberated' approach, pampering the citizenry in order to achieve a golden age there.

Comments?

posted 03-05-99 11:05 PM ET            

For instance, Analyst points out that he can outflank the A.I's build strategy by attacking with a few air units and then overwhelming the happless (and expensive) SAM defender with ground forces. This can be solved by allowing the A.I to view a list of the human player's total forces, and compensate for them as a result (e.g buiding normal garrison to guard against Analyst's wave of ground troops).

Personally, I also find that the A.I makes (very) stupid decisions about where to put it's bases. In fact the first thing I usually have to do is to decide which cities to keep and which to disband - C/o Atrocities R Us (record 320 turns of economic sanctions - man I wish there was a way to turn on those sunspots more often ).

Allowing the A.I to view the immediate region surrounding it (say 15 square radius) *might* allow it to perform this task better. Although factional behaviour also seems to be an influence; the Spartans build their cities VERY close together, while the UoP has them spaced out.

At any rate, there do seem to be ways and means to help the A.I get better at SMAC, even if they aren't really 'legal'. Now, obviously a high quality cheat free A.I would be the best solution, but that doesn't seem to be economicial. And a properly challenging opponent is a nice thing to have (who knows, it might even help to alleviate the pychopathic nature of the A.I on Thinker and Transcend). The point being, is it bad to cheat if it makes the game better?

posted 03-05-99 11:10 PM ET            

Besides, you violate a fundamental law of economics as well: If you increase the price, you geometrically reduce the demand. Consquently, you get more money by lowering the price.

posted 03-07-99 04:41 PM ET            

On the subject of the city engineering screen, it just occured to me that there are a few more uses for it. The other day I was surprised to find one of my spartan cities (Sparta Command) in an almost constant state of revolt (once every two turns). Then I noticed that the population was booming there. I checked the Social engineering screen, finding out I had +4 Growth, a good rating but not enough for a population boom. Must be a bug, I thought.

It turned out that I had simply forgotten about the childrens crech (+2 growth) in Sparta Command. This gave a +6 growth to the city, hence the population boom.

It would be nice to be able to access a single screen for each city which gave all the relevant benefits enabled for that city in a clean, easy to view form. Such as the morale bonus for each type of unit produced there (Social + Infrastructure), a deeper breakdown of the drone/talent situation (i.e distance from capital, number of cities, efficiency), the aforesaid city engineering interface. This could definitely be a useful addition to the city management screen.

Gergi: I only propose enough cheating to allow the A.I to compete fairly under it's own handicaps. Whether problems are fixed either by better A.I algorithms (for instance the appalling use of airpower), or by less elegant methods is irrelevant to me. So long as the game remains challenging.

posted 03-08-99 03:42 PM ET         

BTW, I've noticed that the AI also deploys the bait and switch tactic described by Analyst.

Shining1: Actually, the "poor" base-building pattern on the part of the AI is a deliberate ploy on its part. It reflects its inability to compete directly with you in base "buildup", especially if you have built the right projects (Genome,Vats,TeleTubbieMatrix, etc.). Further, a player actually posted this here as a recommended tactic - just slap them down in a close-knit (~2-3 squares apart) regular grid pattern. Unbeknownst to that poster, the Firaxis AI knows this and does so with a vengeance!

I call it aesthetically atrocious (instant vendetta against factions with ugly base layout!), but the bottom line is that this is more evidence of lack of base / teraforming AI.

posted 03-08-99 04:22 PM ET            

Yeah, the base building algorithm is probably one of the most irritating A.I features. The best method is to save a bit of cash, select the bases to keep, and turn the rest into colony pods (starve them if you have to). If you have the planetary transit system, this can be a quite effective tactic (a population boom also helps).

posted 03-08-99 06:36 PM ET         

Okay, so if need to change your current production item you can double click or better yet hit the 'c' key, nice, convenient. You select your new production item and then you are kicked back to the city screen. Unfortunately, you wanted to queue up a few other items too, no problem you'll just jump back to the change production screen. So how do you do that. Well you have to double click on the queue screen to get back to where you just were. I'm confusing anyone yet, which is kinda the point. If I'm correct I don't think that there is even a hot key to go to the queue selection.

Alright, I'm sure at this point I've lost alot of people. What I'm trying to say is that I believer the user is most likely to use the queue option right after they use the change production option so therefore there should be some intuitive easy way to go from one to the other. In other words you should be able to change your production and queue items all in one screen.

posted 03-08-99 09:57 PM ET         

I would like to comment on Rong's algorithm for directing formers. Rong gives the following as pseudo code for such an algorithm:

1. Look at neighbor squares. If found hostile units (native mindworms, enemy units), move to the opposite direction. End Move.

2. Look at current square. If within my city radius and forest is plantable (not rocky, not xenofungus), plant forest. End Move.

3. Look at neighbor squares. If one is within my city radius and forest is plantable, move there. End Move.

4. If current square is in-between my two cities (i.e. the combined distance from this square to the two cities is less than or equal to the direct distance between those two cities), build road. End Move.

5. If current square is not in-between my two cities (the opposite of 4) and sensor can be build, build sensor. End Move.

6. Look at neighbor squares. If there is one that satisfies 4, move there. End Move.

7. Look at neighbor squares. If there is one that satisfies 5, move there. End Move.

Seems like a straightforward, and potentially successful, algorithm. However, if you'll indulge me, what about the following:

1) If rule #3 is applied, but the former moves to next to an enemy unit or worm? Even if they don't get attacked, next turn rule #1 will apply. Result: potential infinite loop.

2) What if there is no enemy unit in that vicinity, but you 'know' that there may be lurkers over there (because of proximity of fungus or vendatta factions)? Shouldn't a 'safety' heuristic be applied when moving (unless, of course, you have combat units close by that are 'securing' the area).

3) Why build forests arbitrarily everywhere, when you need might be for energy? You might even end up covering energy, mineral, or nutrient bonus squares with forest.

4) Shouldn't some 'greater need' criteria apply after the needs of the city are satisfied? What if our faction has a great energy need (or, more subtly, wants more energy to fulfill a 'grand plan')?

5) What if you only want to build roads directly between cities, and not all over? Shouldn't there be a 'road prioritization' paradigm?

6) What if there is a great need for formers at other cities on the frontier? Shouldn't there be some way of dispatching formers to help out at more critical sites?

Etc etc. I don't mean to criticize the algorithm presented, merely show how difficult it is to get one that is really 'challenging'. As you consider more and more cases, the algorithm becomes harder and harder. Pretty soon, it is so complex and cumbersome that it requires a great deal of resources to evaluate.

Also, one of the great 'dominant' strategies in the Man vs Computer Chess scheme is to find the holes in the AI. Do something stupid, or unexpected, and the AI may crash and burn. It might send formers around mindlessly, or make meaningless and wasteful attacks, etc.

I guess my point is that translating what you intuitively 'know' into a foolproof algorithm is darn difficult. For example, you know not to send formers over there, because last time you did you watched 'em go up in smoke, and you just KNOW the bad 'uns are still lurking out there. What do you do? Wait? Send out some combat units to secure the area? What do you send? What support? And how does that interfere with whatever 'grand plan' you might have?

There are a lot of interesting observations and ideas here, but let's not kid ourselves: this AI topic is tough. Unfortunately, most games today treat the AI as an afterthought, and sink their dev $$ into neato 3D effects. SMAC is one of the few that even attempts to address the need of 'real' gamers: an AI opponent that will challenge us, and not have to cheat (too much) to do so.

The holy grail of gaming is, in fact, a computer opponent that we can all respect.

posted 03-08-99 10:10 PM ET            

As for the size of the algorithm, I doubt it would take more than a couple of hundred lines of code to impliment - maybe a bit more if it was doing everything perfectly. The governor/former combination might take a bit more, since you'd have at least four former behaviours, but then again you can simplify it by including a call to the 'check for hostiles' routine in every behaviour.

posted 03-09-99 01:30 PM ET            

You might also note that, in the game I described, Dierdre was #2 to my #1, yet she remained my pact sister and voted for me for Supreme Leader, so it's not inevitable that #1 and #2 powers will be at odds. Hard to say what the other differences in our games might have been, since you were adopting Morgan's pet social engineering choice.

posted 03-10-99 08:46 PM ET         

When you say that such an algorithm could be written in a few hundred lines of code, I assume you mean pseudo code? I've been slinging code professionally for 20 year, and have fallen many times into the trap between pseudo code and real code. Just consider the following areas that would need to be managed:

1) Resource tracking (where are what resources?)
2) Need management (city short/long term needs, faction short term needs, overall long term 'goals')
3) Threat management. Don't want your formers chewed up by lurkers.
4) Path finding. Want to build best paths for roads (as opposed to brute force 'build roads everywhere')
5) Sensor placement.
6) Local vs 'global' faction needs
7) Green management

and more. I think that each could potentially be boiled down to a set of rules and algorithm, but surely each is a subject in itself.

posted 03-10-99 09:00 PM ET            

I suspect a lot of the complaints about the auto-formers stem from the fact that they work to a plan for the average 'best' city. When people want farms, and the former builds mines all over their best farmland, they naturally say "Duh! That's just stupid!"

If you could get a system to maximise the terrain for either nutrients, minerals, or energy, things would be a lot better.

The key being that the A.I DOESN'T have to decide for itself what the goal is. The rest just flows from the maths of the situation - do the best squares first (including time till city expands, removal of fungus, etc). It would take time, but you could definitely get the algorithm to work.

Okay, including all of that, maybe at least 1000 lines of code is needed, depending on the lanuage (my own humble abilities extend only to turbo pascal - but I'm a quick learner ). Possibly much less.

I dunno. Maybe I can work out a sample (word) algorithm during the weekend, just for the growth option.

posted 03-11-99 08:49 PM ET            

(Incidentally, CTP uses a system very similar to the one you desribe. It will be interesting to see it if works).

As for the rest of your post - I agree, the combat system isn't that great. It needs an overhaul. Although the paper/stone model DOES exist already in SMAC, you just need to look more closely to find it. And the war graphics are something we have all complained about.

But the rest of what you suggest runs contra to the reality of the situation - units DO need bases to be effective in real life (currently, it's unrealistic being able to build things without a specific factory and command centre - but that would impact on gameplay too much). The special abilities, are, IMHO, a brilliant idea. As is the design workshop full stop. I definitely see these two items as non-negotiable for future strategy games - they even have a place in RTS, I think.

posted 03-11-99 10:32 PM ET         

I think you'll find people on both sides of this fence. There are the micromanagers that want more and more detail (see Stars! for MOO taken to the micromanagement extreme). Then, there are those who want little or no resource management (what I call the Command and Conquer cliche; hey, don't get me wrong, I play Red Alert with friends regularly).

In between is Nirvana, in my view. Just enough EASY TO USE micromanagement that allows the control I want, but doesn't interfere with the game. Why? Because I want a STRATEGY game, not a strategy game layered underneath a micromanagement tactical game. I want to see my strategies unfold in a continuous fashion. Say, I am on a campaign; at that time, I want to concentrate on reaching my objectives without the constant interruption of detailed micromanagement. Or, when I'm in 'build up the empire' mode, I want to see steady and timely progress towards that goal.

Too many games offer one or the other; micromanagement that is constantly in your face, or lack of it that interferes with your strategy choices (e.g. Lords of the Realm). What I want is the feeling that I am setting a grand strategy and actually seeing it unfold in a continuous flowing game experience (sorry to get metaphysical here).

How to get that? Micromanagement you can 'set and forget'; that does what you want, not what the designer thought I wanted. For example, why have to set the build queue for each city? Why can't I set a template (or a series of templates), and reuse them (e.g. one for land cities, one for sea cities, etc)? I guess my point was, wait, I'm sure it's here somewhere...ah, there it is; I guess I want grand but not repetitive micromanagement.

posted 03-11-99 11:10 PM ET            

Apparently, the autoformers, and the governor, were intended only for newbies to get the idea of terrforming and base management. So it's no wonder they're bad.

Which begs the question, "What the hell is the A.I doing to be using such poor routines."

Either it was intended that SMAC would go multiplayer much better than it has (also contradicting Brian's views in numerous interviews), or else the strategy people decided that base management was a bit of a non-priority for the A.I.

posted 03-12-99 02:50 AM ET            

I may have missed this complaint in the mountain of (articulate and thoughtful) posts which I just read, but couldn't the AI manage the distances which formers will travel, if not the actions? I remember my first few games of SMAC, in which I experimented with governors and formers. To my amusement (I was winning) my auto-formers seemed to have no compunctions about spending ten years to cross my continent only to build a bunker in the heart of my empire. That curious project completed, said formers would rush back to where they began and begin something equally spurious.
It seems to me that formers should operate in roughly concentric circles from their point of origin.
Also: I appreciate and agree with Analyst's comment that SMAC is won by offense. I also would prefer a game where I did not have to fight my neighbors repeatedly until they were either crushed or subdued. The simple truth is, if you begin the game bordered by either the Spartans or the Believers you need to crush them early - because you are going to have to fight them sooner or later no matter what course of action you take - although this is more true of the believers than of the Spartans.
I'm NOT going to rant about the Believers and my disagreement with the fact that they exist as a faction.
As for combat AI - I'm mostly pleased, but find that it simply cannot deal with mind worms or adapt to them quickly enough - I don't like the fact that PSI-attackers get a three to the defenders two either. I wish the Gaian computer AI would take advantage of their (IMHO) huge advantage in this department - in terms of both combat and scouting unity pods without fear of mind-worm peek-a-boo.
I have a ton of things I'd like to say about SMAC, but they are not-really AI related and should therefore be kept out of this Demon-Boil sized thread.
C'mon admit it, that little militarist in your head says "muah-ahh-ahh, this here Demon-Boil is gonna open me a frosty can o' whoop-ass! every time you manage to get one. It's the best looking unit in the game considering the unit size problems that plague SMAC - every time I see a large icon infantry next to a pathetically small icon infantry a little voice calls out "de plane boss, de plane!"

posted 03-15-99 07:48 PM ET         

The governors of SMAC are all faceless, and have the same skill level (apparently not the sharpest knives in the drawer). Why not enhance this 'defect' into a feature?

Governors have different skill levels. You hire them based on some resume of 'previous' data. They then gain their own track record governing your cities, which you can track. Some get more skillful; others, unfortunately, are not up to the task. Those further from the capital are more susceptible to 'inefficiency' (e.g. graft). You manage governers just like any other unit, even 'disbanding' (e.g. banishment, high tower, or more severe resolutions). You treasure your 'good' governers, and sweat decisions to send them to far away, dangerous cities. You cheer when the lob the heads of the incompetent, and mourn the loss of loyal subjects.

posted 03-18-99 02:42 PM ET            

Analyst, your post brings up an interesting point (again). At what point do the number of possible variables in a strategic situation overwhelm even the best AIs? Adding options to game play does three things: it increases the number of discrete areas an AI has to keep track of, it increases the complexity of the paths the AI has to follow, and it increases the chance of serious loopholes in the CP's strategy.

BTW, I am NOT well versed in AI. I do write software for a living, so I am not COMPLETELY ignorant (just mostly).

One of the problems here is that players will always have more time to come up with strategies then the game developer. There are also many more players than developers, and once one of them comes up with an idea, then the idea is usually available to all of them in a short period of time. This problem gets worse as game complexity increases.

One way to get around this might be to get the AI to learn. There is some limited adaptability in SMAC's AI, but as you pointed out, this too turns into something that is exploitable because it only goes so far. Would it be possible to write an AI for a game like SMAC that would learn new strategies as time goes on?

Even if it were possible, there is still the issue that players collaborate on strategy -- an advantage the AI does not have. There would have to be a centralized server that would receive strategy updates from the gamers' machines, synthesize this learning, and then allow the gamer to download the updated AI module.

To make the system complete, there would have to be some way to rank strategies in terms of utility. For example, terraforming tiles would be a lower utility strategy than using probe teams (VERY simplified example). The difficulty levels would then indicate which quality of strategy the AI would use against you. The top level would basically be no-holds barred.

This would probably be very complicated to do, although it would revolutionize gaming AI if it happened.

I'm curious how the current AI works from a system standpoint. Is it just a single routine that branches down appropriate paths at fixed points? An idea that is slightly simpler than the adaptive AI idea would be to have a modular AI. Each module focuses on a specific, discrete aspect of running the faction. Each module proposes a set of actions to a central controller. The central controller then prioritizes these actions based upon the current game situation and available resources.

A setup like that, if done properly, would be able to respond intelligently to almost any situation. The trick is to figure out how to break out the actions into enough modules. Air, naval and ground combat unit production could merit their own modules. Prioritization of the AI would be dependent upon things like treaty status, faction dominance, % water on the planet, defensive strength of a base relative to possible attackers within range, etc.

posted 03-18-99 04:31 PM ET            

(1) The phrase "it should be easy to come up with a routine that . . ." should be eliminated from this discussion. We can argue about the success of certain design features in SMAC, but we cannot argue that the FIRAXIS guys are smart and that they would have certainly noticed an *easy* solution.

(2) I think that critical comparisons of SMAC's top level against the Civ II Deity level are unfair. If I remember correctly, the computer and human players competed on equal terms at the Prince level (4th from the top) in Civ II, and the AI had huge tech and production advantages at deity. The Transcend level is more analogous to Civ II's Emperor. If you make that comparison, it's clear that the SMAC AI performs much better.

(3) Analyst proposes enhancing AI performance by eliminating decisionmaking steps that do not add to gameplay. (Please correct me if I have summarized incorrectly.) The first suggestion is to eliminate the design workshop. However, I don't believe that unit design is part of the AI's strategic calculus. Every game I've played, the computer generates the same set of units, so I assume that the standard units are just a list. Eliminating the design workshop will not prune that list, so it should not affect AI performance.

(4) I don't believe that "less is more" is a realistic strategy for improving gameplay. The industry has only recently developed commercial software that will give humans a tough battle in the relatively simple game of chess. Therefore, I suspect that strategy games would have to be simplified to something two or three times more complicated than chess before we could get a serious challenge from the AI.

(5) I disagree with the proposition that the addition of diplomacy and AI responsiveness to human player strategies have made the game easier. Yes, there's a response to the AI's tactics. However, it's much more complex and dynamic than in Civ II, where I could overrun enemy cities with impunity by throwing a few diplomats/spies at the wall and then killing the one or two defenders. By way of example, in an early game, I hit the computer with a series of combined rover/air attacks and it responded by putting an SAM unit and an ECM unit in each city. Of course, I answered with mindworms, but had I kept that strategy up indefinitely, I would have had to change as the AI adapted. I think this facility makes the game more challenging.

(6) I disagree with Analyst's view that the game is too complex for multiplay. Although I found the institutional perfidy too depressing, I also flirted with PBEM Diplomacy for a while. Although the mechanics are simple, the actual negotiations are so complex that I found myself devoting two or three hours to each turn. Even so, the game is quite popular. I can't imagine SMAC turns taking that long, so I suspect that PBEM will do well. We'll see.

posted 03-18-99 05:33 PM ET            

On both Civ II Deity level and SMAC Transcend level the AI enjoys a 70% reduction in the costs of everything relative to yourself. That includes the cost of production, research, rushing units, etc. Everything. Identical advantage in both games. Further, given social engineering, these produciton advantages can beome absurdly exaggerated. In a recent game, I saw that Yang had unit production boxes that were only four columns wide! This huge advantage in reducing the size of the base of the pyramid (SMAC is definately a game of pyramiding effect) by 30% right off the bat is the foundation of why this level is hard and it is identical in both games.

In both games, as you move up in levels, the AI's drone (formerly unhappy citizen) difficulties are a "mirror image" of your own, so that the AI expends less effort on drone control and, when you take an AI base, it's drone control feature are always inadequate to your needs.

In both games, my observation (not entirely provable but certainly repeatable) is that the AI "peeks" at your build queues and other behaviors and, to a degree, uses this information to mimic your strategies. If you keep seeing the AI build the same units, Will, that's probably because you keep building the same units. By comparing notes with other players, I've come to see that the AI has a strong tendancy to mimic the unit designs and produciton of the human player during the course of any specific game.

Many players have established, through experimentation, that the AI "knows" where your PB missiles are hiding, even when it has never penetrated your datalinks, and sets out to destroy them with a vengence. You don't get this knowledge for free, but the AI does. That's a cheat. I also think that it gets to "see" your military for more purposes than that. The AI gets a free look (and I think it's a detailed look) at your military capacity whenever negotiating to set it's negotiating mood. Even if that info is never used for any other purpose, well, it's certainly not info you get prior to a conference and, in my world, that's a cheat.

As far as I can tell, the only Civ II cheat denied the SMAC AI is that there is no combat multiplier in the AI's favor in SMAC. It's pretty clear that all of the other old cheats are in force and effect.

That being said, there is some other explanation for why the AI is easier to beat in SMAC at Transcend than in Civ II at Deity. This is an opinion now held by a much larger group of persons than myself. We may argue about why that's true, but not that it is true (at least not realistically from where I sit). I just don't see that the assertion that the SMAC AI doesn't get "cheats" is correct.

To address your comments that SMAC can be popular in PBEM, just as Diplomacy is . . . ("institutional perfidity"--gotta remember that phrase). Consider this. Diplmacy games average about 15 game years total, at 5 "turns" per year, with simultaneous move resoultion, for an average of about 75 "turns" per game. At that, an average game of PBEM Diplomacy lasts about 10 months.

SMAC is a game designed to last 400 turns, with non-simultaneous turn resoultion (each player must take a turn then pass the file), so a game of PBEM SMAC will actually take 2,800 "turns" to complete. Even at the wildly optimistic rate of two turns per day a full length game of SMAC will last about 3-4 years. That's not to even mention how the issue of dropouts, etc. would be handled. not very practical as a solution from where I sit. That get's us back to real time multiplayer which returns us to the tedium issues, which gets us back to the inefficient unit design screen and the unuseable auto formers and governors, etc.

posted 03-18-99 06:13 PM ET            

I maintain that any of the specific issues brought up in this forum could be fixed through normal AI coding. There are areas, such as "building farms in the desert," which really do need to be fixed. Of course, that doesn't bother me, because I micromanage everything.

My comments from earlier are meant to address the concept of an AI being a challenging opponent in general, which I believe was a concurrent line of discussion in this thread. Without an AI that really learns, I think that it would be hard to eliminate all of the loopholes in a game as complex as SMAC for the reasons I specified earlier.

posted 03-18-99 08:03 PM ET         

I guess if you could ask a computer why it is so dumb, the answer would be: You humans get to choose the game. Let's play 'Add', 'Multiple', 'Divide', etc and I'll kick your behind. The point is, there ARE things that computers are good at, and things that they aren't. Actually, to be truthful, there are things that programmers know how to do, and things we don't know how to do. Making an AI that can compete with today's gamers unfortunately falls into the latter category.

So, what can a computer (and its AI) do fairly easily? Some examples:

1) Path finding. I know that there are a lot of dumb pathfinding games out there, but there are also examples and algorithms that can find a better path than you or I.
2) Number crunching. If game results depended on a number of factors that you had to 'figure out', the computer could do that much better than you or I.
3) Redundant tasks. Once a computer can be taught something, it never forgets (unless the disk crashes...)
4) Brute force algorithms. Sort of the same as #2, if there was no 'intuitive' answer to a problem, the computer could potentially just crunch the answer out.

The real problem (ok, one of the real problems) is that the sort of things that the computer is good at are not the sort of things that gamers find to be 'fun'. Also, gamers would consider it unfair to have a game that consisted of problems better solved by a computer; where's the sport in that (like arming the deer with bazookas, suddenly hunting ain't so fun).

I think that if you really wanted a computer AI that could compete, you need to redefine the 'normal' way of looking at the game. Some of the solutions must involve problems where the computer does have an advantage. For example, if finding the shortest path all the sudden became CRUCIAL to a game, then the computer might have a partial leg-up advantage. Redefining the game to give the computer a 'built in' advantage solves the problem in a way that probably costs less than making the computer 'think'.

Of course, if you defined the game so that the computer had an overwhelming advantage, it would not be fun. But dumbing down an AI is not hard at all; e.g. just slow its reaction time down, or throw in some random 'erratic' behaviors.

posted 03-19-99 03:52 PM ET            

Thanks for the reply. I knew the computer cheated at the higher levels, but did not think it worked to the same degree as in Civ II. I think there must be some difference somewhere, because it seems like the two have a different number of levels. (I can't tell because I gave my Civ II CD to a friend and don't have the game on my current computer.)

I don't think there's much AI involved in the standard unit design. I also don't think the computer is copying my designs, as I tend to use a lot of rovers, while the computer uses them relatively infrequently. The computer also uses 'copters more and drop units less than I do. This leads me to believe that there's a list of standard units embedded in the game that the computer always uses and that there's not a lot of processing power devoted to unit design.

I noted that, like in Civ II, the human gets more drones at higher levels, but the number seems different than in Civ II. I'll check that out tonight.

On the question of PBEM, it was my experience that each diplomacy turn encapsulated a number of message interchanges among the players, so that each game phase resulted in about four, five, or six exchanges of mail. I also expect that an MP game would work best with people in different time zones. A four-person game with players in Australia, the West Coast, the East Coast, and Europe could process a turn each day. Finally, the players could get together periodically for a few hours of simultaneous play.

On the topic of former automation, I think that the comments on this thread demonstrate how difficult the problem is. One commenter likes to start with farming everything and then mining. Another builds only forests and roads. One Civ II player on Apolyton wrote that he always had two settlers for each city, which seemed a bit excessive to me.) I like to develop each city differently, depending on the specials and on my strategy for the city. It strikes me that it would be difficult to develop a program to accommodate all of these different ideas of what a "good" automatic former would do.

I think this explains why Firaxis didn't do better. They devoted their resources to enhancing the computer's military and diplomatic skills, which were pretty grim in Civ II, and pretty much left the former and base development skills at the same level as in Civ II. I think that was the right choice. It's such a pleasure to occasionally face a combined assault or a reasonable computer ally.

posted 03-20-99 04:50 PM ET            

I really liked Shining1's governor list. That was much better than the autoformer that does NOTHING important.

I do find the computer less likely to engage in a nuke fest. (maybe I have just been lucky so far) I have painful memories of Chief Sitting Bull lobbing nukes at me every turn to lower the score with pollution when my spaceship was just 5 years out.

To Zied, Strategic Empire sounds like an old PC game I played in 88-90 called Empire: The Wargame of the Century.

I do like the idea someone proposed (sorry, can't read my notes) about general orders like "build mag tube to XXX." Speaking of mag tubes, I find that the AI's do NOT make good use of them. (much like AI's in CIV II did not make use of railroads) I find that a large tactical advantage can be obtained with a "transcontinental railroad" that allows me to reposition my forces in a hurry. Autoformers do not realize this advantage at all.

I have noticed that units are much smarter about staying on roads/tubes. I remember many a time in Civ II a unit trucking along on a railroad only to hop off into the hills because the rails were not straight.

I have not seen a size 15+ city, but I have seen the gaians have several size 14 cities in a row.

I also agree with Chaucer in the fact that all games seem to eventually boil down to a firefight with the 2nd most powerful AI. This gets REAL tiresome. I mean, i am just sitting there with a green economy (as UoP) quietly kicking the butts of the the Believers (who are the enemy of the Gaians btw) only to have Diedre decide now would be a good time to attack my western front. Once she is beaten down to 2 ocean bases, Lal decides he wants a piece of me.

This just gets old. The main way Civ I, Civ II, and, unfortunately, SMAC makes the AI's "better" at higher levels is by making the more belligerent. I am not as good as some of you (like analyst) and I can not see all the AI faults. Besides, I really hold that in the same regard as min/maxing in DnD. Why can't the AI just leave me alone if I leave it alone huh? that is the way the real world seems to work. I mean, when was the last time someone invaded Switzerland?

I really would like it if the computer would just play honest. I mean, why is it that someone I have yet to wrong (or even talk to that much for that matter) suddenly decide that I must be destroyed?

I do find the responsiveness of the AI refreshing, but I also saw Lal kill 4 of his planes because I cut off the direct route to one of his cities and and stupid planes just went back and forth until the died. If the situation were reversed, I would either sacrifice one plane to blow a hole in the wall, or send the planes to a different city.

I hope the flame free environment I have seen remains.

Thanks for your patience.

posted 03-20-99 06:45 PM ET         

The 'airplane running out of fuel' is an interesting example of the difficulties in making the AI 'smart' (or, at least, as dumb as us organic-types). I think there are a couple of basic approaches:

1) Specialization. Make the AI realize this specific situation (or a minor generalization of it) and handle it.

2) Generalization. Make the AI smart enough to figure this out in the general case.

I don't think we're there yet for #2, so to handle this case would probably require #1. Frankly, this situation either didn't come up in the game testing or was determined to be so minor (or happen so infrequently) as to not warrent special code.

This example points out the main reason, in my mind, that gamers see current game AI as 'dumb'. Obvious situations that we would have puzzled out, but were not anticipated by the game designers (or not handled, possibly purposely). Nearly all AI today (especially in consumer games) doesn't have the ability to puzzle out a solution to a problem that was not specifically handled by the programming. If that seems like an obvious statement, read it again. If I tell my kids to put their clothes away, they will hopefully understand that means to take the clothes and put them in their proper place (whether they actually do it is another thing entirely). But even if you taught a program what clothes were, it might intepret 'away' as throwing them away, or just moving them to another location 'away' from here.

Sorry for the simplistic example; I don't want to talk down to anyone, but I think that some of the criticism of the game's 'dumb' AI is missing this point. If the AI didn't handle a case, it is not a reflection of the IQ of the AI; it has none. It is a reflection of the fact that the situation in question was not explicitly anticipated, designed, or programmed (for many reasons: time, speed, playability, etc).

Given all that apologism, here's what I would like to see in a game AI (I'm switching sides, swapping my programmers hat for my gamers beanie). I would really like to see an AI that surprises, fools, or otherwise hoodwinks me. Just once. I don't mean stabbing me in the back; MrFlood already commented on that, and I agree it is tiresome. How about AI that 'baits' me? E.g. puts a scrumptious target within my grasp, whereas if I bite I get swamped from behind. If I get complacent and say "well, I'll just magtube all my forces back to the problem area", have the AI cut off the tubes, isolating a portion of my empire, then taking some key cities. Really unexpected stuff. Of course, I would swear and grunt and reach for the 'Load Game' button to back up, but I would respect the AI. This would make me think harder when plotting my next moves; THATS what I want, to be challenged.

Civ2 was always fun in the first stage, until I was firmly established and had the killer technology needed to mop up (for me, it was always Howitzers). At that stage, I would often just give up that game; what fun is stomping a helpless opponent? It just became an exercise in logistics.

posted 03-22-99 04:05 PM ET         

IF nearest city is starving
THEN build only farms

If Firaxis do use approach, it would be feasible to design a game where the player could provide his or her own algorithms for unit behaviour.

posted 03-22-99 05:56 PM ET            

A few points I'd like to make on the AI.

First off, AI programming is not easy. In its current state, AI is just a dressed-up name for complex algorithms. Sure you can slap together 100 lines of code for an auto-former, but how long would it take to execute? Here is what I mean. In one of my games I was trying to build mag tubes to link my cities together, but I didn't want to over build since once a city is linked you can always get there at zero cost. So how do I build minimal length of tubes to link all the cities? Then it occured to me, ha, this is exactly the Travelling Salesman Problem. If you didn't know (you CS majors bear with me), the TSP is like this: say you have a bunch of cities, e.g. New York, Chicago, Dallas and L.A., with different cost (distance) associated with each city pair. A salesman, starting from New York, wants to visit all the cities once while travelling the minimal total distance. Simple you say? Well, the TSP happens to belong to a category of problem called NP-Complete, which means no one has ever (and most likely never will) found a polynomial solution to any of the problem in the category. The significance of that? Say you have 30 bases, and something takes n^2 (polynomial) to compute, you'd need 30^2 = 900 computations. For an NP-Complete problem, like TSP or building mag tubes, the cost is exponential. For example, 2^n would take 2^30 = 1073741824 computations! You want it to terraform or you want it to sit there and think? Forever?

This is only a simple example. You can see a "simple" math problem can turn VERY expensive computationally. In most cases we cannot offord the optimal answer; the best we can hope is coming up with some heuristics for some "pretty good" answer. But how do you come up with good heuristics? How good is good? Things get hairy there.

I guess my point is, don't blame everything on Firaxis (well, they do need to improve their AI at places). Blame your puny PIII 500Mhz for not up to the task of intelligent computing.