H5 AI harder than H3 AI?

[Yurian Stonebow]

"H5 AI harder than H3 AI?"


Is that a trick question? I play H5 with a friend of mine, (patch 3.1 for the Tribes of the East) and so far have seen little else than that age-old "rushing tactics" in the early phases of the scenario. Both multiplayer maps and the singleplayer ones seem to be like this. I highly doubt if the AI in H5 is even capable of adjusting itself to face allied human players or a single human player. Rush, rush, rush all day long. Rush, rush, rush while he sings his song. I watch the AI, gonna make me smile. Will repell his attacks with the modest experience of mine.



Yurian

[Pol]

Never played it, but apparently Galactic Civilizations II (a turn-based game ala Civilization IV and Heroes) had some really well-written AI, capable of holding their own and even outplaying human players. I don't see any reason why the H5 AI can't do the same

Let me know when Nival's CEO can write the AI himself. That's what it takes.

[Asheera]

@Yurian: At what difficulty? At Heroic the rushes are pretty hard to counter, especially the first one (because you start with a few resources)

[Yurian Stonebow]

@Asheera: Good heavens, we're still both kinda new to Heroes V, playing at Heroic would kill us. Remember that I said that "with modest experience of mine". The difference is that I consider this friend of mine as a mentor to me, not an equal. I've learned by watching him play and listening to his advice. The guy has beaten nearly whole of H5 Standard and some of the campaigns of the Farmers of Hate. He plays most likely on Normal/Hard while I use (still) Normal as the level of difficulty when playing alone.

When we play together in Allied mode, he's the one choosing the level of difficulty and giving (endless, lol) advice what to do next. As you can gather, we are but a Novice and an ordinary master, not grand master. We try our best to enjoy H5 but unfortunately, we remain both unimpressed by Nival's handling of the series and the quality of the AI.
Let us hope that the AI will see one final fine-tuning in the long since promised patch 3.2. Well, I doubt this myself...

I can speak only of my experiences (and that of my friend), but the AI at Hard seems to offer only a bit more resistance. The AI of H3 was, in my most humble opinion, a more sophisticated one and in better control of its duties than that of the H5 one. One could easily argue that the H3 AI is more fine-tuned than its "successors".


My two €uro Cents.



Yurian

[Asheera]

Ok people how about we discuss the combat AI as well? I'll throw my two cents here: I think the H5 one is pretty good. Not excellent, but good: it knows a lot of tricks using creature abilities and such (for example, knows how to avoid Fire Breath)

[Kristo]

I've always considered the battle AI to be a solved problem. It's basically chess with a (much) higher branching factor. The computer's ability to fight battles is limited by how much processor power the devs want to devote to it. And how long the human player is willing to wait for the AI to make its move.

Then again, I could be completely off my rocker. I've seen 5 generations of battle AI all do silly things from time to time. I have to believe that people who get paid to write computer opponents would've thought of chess. I wonder just how high the branching factor really is.

[Rife]

In terms of battle AI, in the contricts of Heroes I think it comes down to how well the computer uses the spells and abilities of the factions. H5 doesn't do badly here but I can't say it has ever surprised me. The only battle AI in Heroes where I've ever been caught off guard is in H4.

[Asheera]

Oh and btw, just one tip for those that play below Hard: the AI is handicapped intentional in combat at Normal and Easy, so don't judge it on those difficulty levels. For example, it doesn't cast 4th and 5th level spells, always centers AoE spells on units, etc.

[PhoenixReborn]

I can't say that the h5 ai combat is any good at all. I usually play on hard difficulty. The enemy will cast slow instead of mass slow. It will cast (not mass) deflect missile when I have a dungeon army with no shooters.

I still do it because I need to train my creeping.

[Asheera]

That's weird, I never encountered such silliness before.

[PhoenixReborn]

I also saw an enemy hero cast a direct destructive spell and resurrect my magnetic golems once.

[parcaleste]

^ Maybe the computor is not aware of the Magnetic's special.

[ThunderTitan]

I've always considered the battle AI to be a solved problem. It's basically chess with a (much) higher branching factor.

Yeah... problem is that chess is a lot more linear then you think... move patterns are predetermined and it's easier to give each piece a value... thus the math formula is more efficient.

Oh, and Deep Blue cheated anyway, as it was updated during a break to account for stuff the human has doing...

[Kristo]

In chess, you have your choice among many pieces to move, with a few possibilities for each piece. In Heroes, you don't get to choose who moves next, but there are lots more possible move locations. My gut feeling is that this more or less balances out. As for scoring any potential state of the battle, is HP not accurate enough? I mean, the point of a battle is to reduce your enemy's total HP on the board to zero while preserving your own as much as possible.

Chess AI's are scripted with opening books. There are predetermined good sets of moves you can make in the early game, each with its own counter-strategy. That greatly simplifies how much work the AI has to do early on and it's something to think about for a Heroes battle. Can we define an opening book for a Heroes battle? Can an AI know that if a certain special ability is on the board, it has to act to counter it in one specific way?

The other key difference between Heroes and chess is that most Heroes battles are not fair fights. It's not necessary to write a grandmaster AI when it's going to lose most of the time anyway. We just need it to act sensibly and not make silly mistakes.

[ThunderTitan]

Nope, those choices are less important... in chess it's easier to predict where pieces can move because the move patterns... while in HoMM you can even not move at all... so it's easier for the computer to "see" all possible moves and be programmed to respond to them in chess.

Not to mention that programming chess AI has been done for decades, and no video game will ever have an AI customized for it for such a long time.

going from the earliest real attempts to programs which challenge the best human players took less than fifty years.

Sure, the core concepts are the same, so that's an advantage when programming the AI for other games then chess, but you still have to customize it for the rules of that game...

[ywhtptgtfo]

In addition, there's a difference in the degrees of freedom.

[Kristo]

Can you guys explain what you mean by "move patterns" and "degrees of freedom"?

[Asheera]

Chess is A LOT simpler than HoMM 5. No random factors, no special abilities, no attributes (Initiative, attack, defense, etc etc) or some other variables modifying your creatures. And a lot smaller board. Programming a chess AI is at least ten times easier than a HoMM 5 one.

[Kristo]

With the exception of the board size, all of those variables do not appreciably affect the algorithmic complexity of a computerized opponent. Chess AI's play by brute force. They evaluate every possible move, followed by every possible countermove, followed by... etc., as far ahead as computing power and human player patience will allow. I'm positing that a Heroes battle AI can be written in the same way. All of the variables (stats, specials, randomness) merely affect the amount of work you do when scoring what happens when one creature attacks another. Simple expected value calculations would suffice.

The size of the board and creature movements affect the number of possible moves for each turn. That grows the runtime exponentially but it doesn't make the algorithm any harder to write. I'm arguing that computers have advanced far enough to create a reasonably effective AI opponent in this fashion.

[Asheera]

That 'database trick' with every possible move doesn't count, because I don't think any TBS game has such a thing. You should try and make the AI smart without any database at all based on all the factors there could be in a given situation. It is extremely hard, trust me, that's why people started to use a library database for chess AIs.

It is already hard for chess to do a complex and smart AI without a library database, and a H5 one would be a lot harder than that.