The AI... when will ìt earn its "I"?

[geoschmo]

I think some people generally think that this kind of programming then it easier then it really is.

Consider for example the programming needed to create a quality chess playing AI program. There you have a game with a much simpiler rule set. Two players, 16 pieces of only 5 different kinds on an 8 by 8 grid. Most pieces have only 2 or 3 rules determining everything they are capable of doing in the game. Seems like it would be easy no? In fact programmers thought it would be easy. But it took decades of work by hundreds if not thousands of programmers world wide to come up with a program capable of beating even a moderatly skilled player. It was only in the Last few years that one could defeat the best players, and even now it doesn't beat them every time.

Se4 by comparison is much more complex. Exponentially more variables, just in the stock game, not even taking into consideration what can be done with mods. It's not suprising that the AI is much more difficult to program. It's actually suprising that it's as decent as it is.

It's no suprise that the game industry has gone towards real time strategy. With real time strategy it's much easier to give a human player a tough game because humans can't think as fast as computers can. This does wonders for masking repetative or illogical behavior by the AI. With a turn based strategy game the human player can think about their moves as long as they want before comitting to them. Remeber, computers, even the big ones, aren't smarter then humans, just faster. Make them wait on us the computer doesn't stand a real chance then.

Geoschmo

[ December 21, 2003, 23:37: Message edited by: geoschmo ]

[JurijD]

Well I have some experience in programming AI and I have to say that there are 2 basic strategies in providing a chellenging AI for computer games.

1. You create several strategies for the AI based on what humas do and let it combine those strategies in a random way hoping it will make something smart.

2. You try and pre-program almost all possible eventualities in a game and the "AIs" responces.

For a game such as SE IV the second option is not even to be considered and the first one is mostly implemented. BUT the major drawback is that the AI doesn't learn. The biggest problem in creating an AI isn't in programming its tactics and strategies... those can be simply random combinations of several hundred pre-programmed ones.. the biggest challenge is making some kind of algorithm that EVALUATES the EFFECTIVENESS of the strategy chosen and makes changes to it in order to achieve better results.

Good results in this area have been made in the field of genetic algorithms and neural nets but the thing is that these approaches are not developed enough to take on the wast scope of decisions needed in SE IV. People tend to underestimate the human brain when it comes to AI. Remember we have something like 10^13 neurons in our brain and more than 10^16 connections between them... thats a lot of "adaptable code". And we had a couple of billion years to get where we are now. It would be silly to expect us to pre-program something even remotely so versatile on a couple hundred megs of CD space...

I think that the future of AI in games such as SE IV is to get the AI on aditional CDs apart from the Game itself. The AI could be (in theory) developed if you put several million adapatable and self-changing genetic algorithms to compete against each other on many millions of games... then eventually you'd get something that is worth comparing to humans. But noone can really answer how long this process would take or how powerful the computers would have to be. But for now its the only way as we are not intelligent enough to write the program ourselves... but with some luck and enough processing power we could make it happen randomly.

[Roanon]

I agree, Geo, but only partially. Being a programmer myself, I know how difficult it would be to programm a real AI. To be as creative as a human, as resourceful, considering all options and possible strategies derived from whole sets of options, thinking ahead a few turns and choose the best way to act for multiple units as a strong, simultaneously striking force, persuing a final goal consisting of numerous single targets and objectives - impossible with the technology at hands right now.

AI able to ACT like a human - ok, lets put that aside. But what I, too, consider as "not doing a job" is really *stupid* AI that does not even know how to REACT properly. There are a lot of simple problems - own fleet out of supply, enemy fleet near own planet, enemy planet in friendly system, etc - that the current AI has no real Algorithm for to cope with. Instead, it seems that each turn each fleet decides anew what to do, independent of all other fleet actions, unaware of longtime major problems and the overall global situation but reacting to nearby local events happening Last turn, mixed with a good shot of the random generator. This could have been done better in nearly all games I know. Though this also is not simple - for small companies like MM it might just be too expensive, but I cannot accept excuses like this from big companies obviously spending like 100 times more money on graphics and design than on AI programming.

Yes, it might be possible to play a challenging game vs. the AI. But this is always happening like an elderly grandpa engaging in hand-to-hand combat against a young strong brutish bully - and winning finally. Looks challenging, is not. And advices like "if you want it more interesting, tie the grandpa one hand on the back and blindfold him" may sound interesting, but he still is winning, and this just makes it more ridiculous.

Thank god (and Geo!) for PBW !!!

[jowe01]

When I started playing SE4 three years ago, I found that it was a game with great concept, mechanics and customizability. However, the AI was a joke then. A couple of weeks later, Aaron made a significant improvement to AI agressiveness, which made the game playable in single player.

But since then... nothing. Or very little. In single player, your adversaries are still if not braindead, then at least seriously deranged when it comes to their military campaigns. Aaron and the modders did a very good job improving AI resource management and building programs. Mods like "AI campaign" have taken this a step further, giving the AI "unfair" advantages to further improve its strength. However, this approach does not yet solve the fundamental problem of the SE4 AI. It is like an already very muscular bully who gets some more anabolics. Beefed up like this, he comes running at you and you think "oh,my, it's over now for me", but 3 meters in front of you, he tuns around and runs somwhere else.

This is excatly the behaviour of the SE4 AI. It just does not follow through. In my current "AI campaign" game, again and again I had 30+ Jreaner ships entering my border systems which were virtually undefended. I think "oh,my, it's over now for me", but 3 squares in front of my planet, his fleet turns around and leaves through the next wormhole for somwhere else. This happens again and again. Basically, if you are not extremely unlucky with your starting position, the AI never offers a challenge.

I cannot imagine that it would be so hard to change this behaviour. Basically you would have to somehow drastically increase the priority of an ennemy target, once it has been acquired (thus only very, very urgent new threats would have the AI abandon its previous targets). The current AI behaviours largely decreases gaming pleasure in single player. I imagine, that it is probably to late to expect Aaron to again tackle the AI in SE4, but I just hope, it will be significantly improved in SE5.

Anybody any thoughts on this (same experience, different experience, infos on SE5 AI, ...)

[geoschmo]

Originally posted by Roanon:
Instead, it seems that each turn each fleet decides anew what to do, independent of all other fleet actions, unaware of longtime major problems and the overall global situation but reacting to nearby local events happening Last turn, mixed with a good shot of the random generator.

Actually, as far as Se4 is concerned, it not only seems that way, that's exactly the way it is, except for the random generator part.

[ December 22, 2003, 00:19: Message edited by: geoschmo ]

[DavidG]

just out of curiosity can anyone name a turn based strategy game with a good AI?
(AI's that can win by just boosting their resources, research etc way higher than the human player don't count)

Of course there are a few simple things that could be done to improve the SE4 AI. One I can think of is for the AI to recognize when it runs into a minefield and not do it 20 turns in a row.

[Gryphin]

My thoughts on AI
AI = Algorithm Intellect

The issue behind Algorithm Intellect is the amount of time and money it does take to code it. You mentioned one situation that has occurred. We have all seen bizarre behavior on behalf of the Algorithm Intellect in SEIV. How many lines of code would it take to cover that one situation? How much are you willing to pay for it? I feel the game is a very fair price for what it provides. I fear an Algorithm Intellect able to consistently challenge a human would cost 1000s of dollars.

Hope this helps.

[Fyron]

Galactic Civilizations has a fairly good AI.

[Starhawk]

Try playing the AI on high and in a big universe where the Comps have enough time to build their fleets like you do. And start at absolute low tech and with the medium resources and the like. That should give you a challenge because it has given me a challenge and an enjoyable experiance in almost every demo game I've played of SEIV.

Of course some races seem to be stupid like for example the EEE which always underarm their ships compared to EVERYONE else as well as have inferior shields and armor to just about every other race I've ever seen play against them.

I mean this one time I saw a fleet of Jreanor controlled by the AI attack a LARGER fleet of EEE ships and come out victorious!!!

I mean is it just me who thinks the EEE are mentally slow when it comes to arming their warships for war and not just a little girl slapfight?

But there are things like TDM which I am not quite sure if all it does is improve the AI or if it screws around with absolutely everything. Could someone please tell me exactly how TDM works, does it just improve the AI or does it as I stated above screw with building and science and other stuff like that?

[Fyron]

TDM is an AI mod. It does not change the stock data.