Will the ToE AI be less dumb??

[Alamar]

Anything but "trivial" AI fixes takes a HUGE amount of time and I'm not sure if UBI-Nival has enough $ / time / resources / know-how to give us the fixes that we'd like to see.

I'm not sure if the company's organization would even allow for this considering that it would be hard to justify the ROI to the bean counters.

------------------------------------------------------------

I think our best hope if to convince UBI and H6's developer that the only acceptable AI is "the best HoMM AI period". Considering the advances in CPU / memory / multi-threading / etc. I wouldn't think that the above goal would be that hard to meet ....

--------------------------------------------------------------

As far as H5 & memory goes I wouldn't want to play with anything less than 1.5GB ram. 1GB of ram on a squeaky clean XP load may be enough but that's pushing it ....

[MistWeaver]

Considering the advances in CPU / memory / multi-threading / etc. I wouldn't think that the above goal would be that hard to meet ....

As Ive said before, in HoMM AI development you cant rely most of it on calculating power, its not a chess AI.

I think our best hope if to convince UBI and H6's developer that the only acceptable AI is "the best HoMM AI period".

Agree.

[Alamar]

Considering the advances in CPU / memory / multi-threading / etc. I wouldn't think that the above goal would be that hard to meet ....

As Ive said before, in HoMM AI development you cant rely most of it on calculating power, its not a chess AI.

I totally agree that HoMM can't use a brute force chess AI.

I still think though that if H6 doesn't have the "best HoMM AI ever" then it's a failure ....

[Kristo]

I think they should start with an AI that has to play by the same rules as a human player. Then they can worry about making it the best ever.

[Jolly Joker]

That is pretty ridiculous. It's just not possible to make an AI that will be a challenge for a halfway decent playing human without the AI cheating, because then it would be possible to find winning strategies with simple if-then and either-or directions.
However, to NAME winning strategies you must KNOW them first - you must KNOW what is good play (and what is bad) because otherwise you can't program the AI to do it.
So for all of you playing the game now for a year or so, imagine this little task. Your older brother is a very good player (and you are good, too), and he likes a real challenge you can give him. However, you'll be away half a year and you are owing your elder brother a challenge. You have a latecomer as a brother, 4 years old, no idea about anything, but he listens, when someone talks to him, and he can memorize everything.
Question: Can you tell him how to play, on any map with any race in any situation, to be a challenge for your elder bro? Remember, only simple directions are allowed, nothing abstract. No "judgement", no "common sense".

So very obviously, the more interesting the game is in terms of decisions to be made, multiple roads to walk on, in short, the more complex and intricate the game is, the more difficult it is to build a decent AI.

The only alternative is making a SPECIFIC AI: always develop so and so, build your hero this specific way (and of course cheat, if something is missing like resources, money or choices).

So trying to develop an AI that plays under the same rules than the humans would be the wrongest way to go at it, simply because it's an impossible way (and if it was possible the game would suck). No, the aim is - very obviously - to make the game as easy as possible for the AI because the AI can't be creative.

[Gaidal Cain]

I don't disagree with the notion that the AI has to cheat, but there are better ways to do it than what's used now. The AI is there to pose a challenge, and preferably one that's a s close to what you'd get by a human player as possible. To do this, the easiest way for me seems to bend the rules for it, not break it: give it a bonus on all it's income (+50%), not lump sums each week, in order to make it possible for a human to deprive it of needed resources. Make it take fewer losses than quick combat would indicate. Give it a larger sight radius etc. This sort of thing is no harder to program than any other kind of bonuses, but it will help give an illusion of the AI following the rules. The AI will have to cheat, but the less blatantly it does so, the easier it will be to both control for a mapmaker and the less frustrating will it be for the player who wants to try another strategy that would work against someone playing "by the rules" but can't work against the AI since it's been made immune.

[Jolly Joker]

That's just playing with words.
I see this much simpler: if a human plays on heroic this would imply that the opponent AI would play on easy (not on heroic as well). This would mean, the AI would start with, what, 40000 gold and assorted resources. Furthermore this could imply - legally - all kinds of bonusses: an already built resources silo; one or two additional and already flagged gold mines and so on. Furthermore, for battles this would imply reduced neutral stack strength that would give the same XP and so on.
In other words, it's perfectly sensible that under hard and heroic difficulty the AI plays a different game, a game under easy and normal difficulty, respectively and for example.

[Silence]

A small though about the AI.

In real time strategies the AI challenge is no problem, since AI is simply faster in handling its stuff than human. The lack of intelligence is compensated with the speed of decision making, micro+macromanaging.
Is anything like that possible for a turn based game? Theoretically not, but practically perhaps. The AI indeed cannot outspeed the human anymore, yet it is still better in handling large datasets. A very open field of decisions can be more restictive to human than to AI. Simplest illustrative example is automathic pathfinding between two points crossing sections of roads and deserts: the AI has no problem finding the fastest way costing least movement points. But an average player would not bother with precise calculations and make a move that *feels* or *looks* the fastest. Hence, wouldn't open maps with a large number of towns and requiring good logistics and managment of many heroes instead/beside one superhero be a way to balance AI incompetence? A map where each player has just one town and needs one or two heroes is more a RPG map where town and hero buildup are more important than handling a large set of possibly decisive parameters.

[Gaidal Cain]

That's just playing with words.
I see this much simpler: if a human plays on heroic this would imply that the opponent AI would play on easy (not on heroic as well). This would mean, the AI would start with, what, 40000 gold and assorted resources. Furthermore this could imply - legally - all kinds of bonusses: an already built resources silo; one or two additional and already flagged gold mines and so on. Furthermore, for battles this would imply reduced neutral stack strength that would give the same XP and so on.
In other words, it's perfectly sensible that under hard and heroic difficulty the AI plays a different game, a game under easy and normal difficulty, respectively and for example.

No- that's playing the same game. If the human would come and take the gold mine away, the AI would be hurt by it. The AI should be forced to care about what happens on the adventure map. An AI that's given all the money and resources it wants won't be a good challenge strategically, as there's no need for it to be programmed to care about it. It will only need to manage tactical combat to be able to win the map if it can show some agressiveness.

[Romanov77]

Guys....you are going a bit off topic...


My request isnt about having a super-smart AI that can compete with human brains......
I think that a Homm3 like AI would be already perfect (man, homm3 adventure map AI was wicked)


The problem is the following:

Will the ToE AI take DECENT times to make their turn and avoid retarded behavior like going back and forth on the same 2 squares?

[Jolly Joker]

That's just playing with words.
I see this much simpler: if a human plays on heroic this would imply that the opponent AI would play on easy (not on heroic as well). This would mean, the AI would start with, what, 40000 gold and assorted resources. Furthermore this could imply - legally - all kinds of bonusses: an already built resources silo; one or two additional and already flagged gold mines and so on. Furthermore, for battles this would imply reduced neutral stack strength that would give the same XP and so on.
In other words, it's perfectly sensible that under hard and heroic difficulty the AI plays a different game, a game under easy and normal difficulty, respectively and for example.

No- that's playing the same game. If the human would come and take the gold mine away, the AI would be hurt by it. The AI should be forced to care about what happens on the adventure map. An AI that's given all the money and resources it wants won't be a good challenge strategically, as there's no need for it to be programmed to care about it. It will only need to manage tactical combat to be able to win the map if it can show some agressiveness.

Yeah, but by the time the human COULD take that gold mine away it would already be too late - and the gold mine COULD be unaccessible, not an unheard of feature in single-player maps to feed the AI with gold.

The bottom line is - for me - that the AI is NECESSARILY playing another game and therefore the strategies needed to beat it differ from beating a human. It might get money out of the blue, but it doesn't get creatures out of the blue (which was different in H 2, for example, where the AI would get more creatures per town, if I remember right).

See it from another point of view. If you play on a 4-player map with 4 humans there is no way to win a simple destroy-them-all map, if no one blunders massively or tries to overcome the stalemate (with someone else profiting from it). You do want to play those maps, though, so the AI must necessarily be limited otherwise there would be no way to win, wouldn't it? I mean, if you play that map and attack one opponent, do you really stand a chance when another opponent just picks that moment to attack YOU?

So the AI must play another game than a human, because indeed its goal is not to win, it's goal is to make the game playable alone.

[ThunderTitan]

That's what i always loved about most AI's, especially in missions, if they were playing the same game you wouldn't have had a chance.

But you know what JJ, just because that is required in the campaigns or a story map shouldn't mean it's always like that. 4 allied AI's should crush you on non-story maps where you start equally unless you're very good/lucky. That way you can learn to play better without human opponents.

[Gaidal Cain]

Yeah, but by the time the human COULD take that gold mine away it would already be too late - and the gold mine COULD be unaccessible, not an unheard of feature in single-player maps to feed the AI with gold.

Yes, but that's another thing - that's the mapmaker trying to make up for deficits in the AI. And it's a singleplayer map, which means the challenge has to be different. The problem would be if you from the idea that the AI needs to be fed gold on singleplayer maps infer that it always need to be done so, you might end up breaking some balanced MP-type map. The Ai should be made as good as possible, and then there should be bonuses added depending on the map.

It might get money out of the blue, but it doesn't get creatures out of the blue (which was different in H 2, for example, where the AI would get more creatures per town, if I remember right).

The problem with that is on some maps, money is the limiting factor, not growth. The AI should be made so it plays well on any type of MP map, not so that it's unbeatable on really really poor maps (because of static bonuses) and worthless on really rich ones (because those bonuses are fitted to normal maps and not enough on richer). Once again: the AI can get all sorts of bonuses, but these should adjusted to what's going on on the map.

See it from another point of view. If you play on a 4-player map with 4 humans there is no way to win a simple destroy-them-all map, if no one blunders massively or tries to overcome the stalemate (with someone else profiting from it). You do want to play those maps, though, so the AI must necessarily be limited otherwise there would be no way to win, wouldn't it? I mean, if you play that map and attack one opponent, do you really stand a chance when another opponent just picks that moment to attack YOU?

I think it's quite silly to think that more opponents means a more gratifying game. I'd rather have an AI that challenges me through "brilliance" than sheer numbers. If that means that 4-player maps aren't feasible, so be it. There's no fun in starting a game if you're absolutely sure you're going to win beforehand ("win" being defined broadly - it might be about setting a new record or something).

[Jolly Joker]

@ TT
Sure, I see it before me, players losing each map, and being happy about it: "That's what I call an AI, I never won a game against it, but I learned so much for playing online. A pity I don't like playing online due to having to wait on your opponents."

[Silence]

@ TT
Sure, I see it before me, players losing each map, and being happy about it:

Take H3 small sized random map with 7 allied AIs against yourself. You can win it due to luck, but if you expect the AI to be there just to give you a more pleasant win it is certainly an idiotic thought.
The AI has to be as good as it can and up for a victory not for "just filling empty space". Haven't yourself said that no matter how good the AI is, it cannot win a creative mind? The only way to make the game unwinnable for human is exactly to make the AI play by completely other rules and therefore immune to the human strategies.

[Silence]

Guys....you are going a bit off topic...


My request isnt about having a super-smart AI that can compete with human brains......
I think that a Homm3 like AI would be already perfect (man, homm3 adventure map AI was wicked)

It would be "off topic" or no topic i think. They wont use H3 AI. IIRC they had a look at NWC works (including the NWC started H5 project) but found it easier to start from scratch than learn they way of thinking of NWC scripters and try to mend it with the new concepts.

[Jolly Joker]

No, that's silly. The AI is playing a different game to make it able to compete at all. Of course, if you play a small map against 7 ALLIED AIs you'll of course lose. However, if you play against 3 human opponents the opponents don't have to be ALLIED for you to have an overwhelming chance NOT to win (a simple destroy-them-all). So if you play that same map against 3 AI opponents you should have a BETTER chance to win because otherwise it would be no fun to play those maps.

On the other hand the AI has to be able to play good enough to make it difficult to win (so that you will lose occasionally). The bottom line is that the AI is playing a completely different game. If you can't see that, you should put some consideration into it.
1) The AI can react on human moves only as far as it is programmed to do so
2) The AI can develop a map-specific strategy only as far as it is programmed to do so; does someone really think the AI programmers would be able to foresee all possible map constellations that may occur on any map.
3) In developing "strength" the AI can be only as good as the programmers were advised what "strength" really means. What strategies and moves are better than others. In other words to be able to program a good AI you must be a brillant player.
You can check that for yourself. If a chess-AI is evaluating moves it must have an "idea" of HOW to evaluate (what exactly makes one move better than another one).
The same is true for Heroes, of course. Is it better to invest 5 Crystals into an upgrade to be able to take out some vital stack or is it better to wait and go for the next level dwelling? Answer: IT DEPENDS ON THE MAP SITUATION!

So, there is NO way whatsoever to make the AI play a decent game without cheating, i.e. letting the AI play a more or less different game.
IF you do that - and there's no way around it - you obviously can make the AI invincible with cheats: give them additional troops, advantages in battle, better luck modifiers, unlimited money, level 30 heroes, whatever. So the actual art in programming an AI - on the difficult levels - is supplying a hard, but still beatable opponent.
The real problem is the lower difficulty level, though. Because here play is easier for humans even without the AI, however, the AI must not be overwhelming - but should be a challenge nonetheless. A real problem and one that H 5 certainly didn't solve.

[Wolfsburg]

100% agreed, JJ.

[Silence]

However, if you play against 3 human opponents the opponents don't have to be ALLIED for you to have an overwhelming chance NOT to win (a simple destroy-them-all). So if you play that same map against 3 AI opponents you should have a BETTER chance to win because otherwise it would be no fun to play those maps.

You HAVE the better chance to win against AIs, EXCEPT when when the AI game is way different from human player. And under different play I mean the general rules how resources, armies etc are gathered, how logistics work an so forth, not that one prioritices defensive game other offensive or one makes decisions based on algorithms other based on experiences.

So, there is NO way whatsoever to make the AI play a decent game without cheating,

As I posted on last page, I am not convinced in that. Theoretically yes, in turn based game there is no way whatsoever as you say, but practice is IMO something else than theory. Creativeness is a weakness of computers, but an average human does not handle large data quantities and do calculations accurately. To have a competitive AI, the game does not have to make it cheat, but the game may be done so that complex computations (logistics a.la chaining for example) where AI easily beats human is a significant part of strategies.

[Kristo]

1) The AI can react on human moves only as far as it is programmed to do so
2) The AI can develop a map-specific strategy only as far as it is programmed to do so; does someone really think the AI programmers would be able to foresee all possible map constellations that may occur on any map.
3) In developing "strength" the AI can be only as good as the programmers were advised what "strength" really means. What strategies and moves are better than others. In other words to be able to program a good AI you must be a brillant player.

So, there is NO way whatsoever to make the AI play a decent game without cheating, i.e. letting the AI play a more or less different game.

That conclusion doesn't follow from your three premises above. It *is* possible because it's been done. Brad Wardell, lead AI developer for Galactic Civilizations II, has lost to his own AI more than once. It doesn't cheat unless you allow it to (via game settings). He's played hundreds, if not thousands, of games against it, constantly updating and adding algorithms to counter specific strategies. After one of the recent patch releases he stated that playing against his AI is basically equivalent to playing against him. Yes, it takes year(s). Yes, it takes a brilliant player to develop it. But it's been done, thus proving it is possible.