Recommend
2 
 Thumb up
 Hide
5 Posts

Through the Ages» Forums » General

Subject: More information about the AI? rss

Your Tags: Add tags
Popular Tags: [View All]
Alexander Stumpp
Germany
flag msg tools
I've worked on Ai for Pandora First Contact, Dominus Galaxia and have written a Bot for Starcraft using Bwapi. So you could say I'm an Ai junky and am very interested in that regard.

It is very rare that you see an Ai in a game that is actually competitive and offers a decent challenge to human players without the need of bonuses. But recently I stumbled upon such a game: The through the ages App!

I have a hard time beating medium and haven't even played against the highest level. And this is rather fascinating. So I'd like to read up a little more about how the Ai works.

I read that it uses Monte Carlo tree search and a rather sophisticated evaluation function.

I'd like to know if the tree search uses pruning of obviously bad moves to score a higher search depth and how deep it usually gets.

Another question would be whether the weighing of different resources is also subject to some kind of randomization or even could be tied to the color of the opponent Ai. I ask because I have seen vastly different focus from the Ai. Is the difficulty changing the weighing always in the same way or does it just increase an already existing variance of the randomness to also include rather illogical weighings?

I would really like to read the actual code. I suppose the weighings are mostly handcrafted as in traditional non-alpha-zero chess-engines.

Not that I have a clue about neural networks, but whenever humans can't describe exactly why a decision they make is the best and they say it's "intuition" is where NNs really shine and outperform us with the intuition they built by playing millions of games.

So if anyone has more information about the AI and be it just theories, I'd really like to read it!
1 
 Thumb up
 tip
 Hide
  • [+] Dice rolls
Pieter
Netherlands
Tilburg
flag msg tools
Good intentions are no substitute for a good education.
badge
I take my fun very seriously.
Avatar
mbmbmbmbmb
Monte-Carlo Tree Search has no limited depth. It searches until the end of the game, using random moves. The evaluation function ensures that those random moves have a higher probability to search the stronger rollouts.

For MCTS, you have to ask how many rollouts it gets to play per move. But more interesting is what the evaluation function is actually considering.
1 
 Thumb up
 tip
 Hide
  • [+] Dice rolls
Riku Koskinen
Finland
flag msg tools
mbmbmbmbmb
Different AI focus seems to relate to what cards they already have picked and what military cards they've drawn.

Sadly even the expert AI makes moves sometimes that don't make sense, i.e. they leave a CA unused with plenty of hand space and there being a useful yellow card to pick. Or they switch to a new leader from Aristotle and then pick a tech card (with available hand space before leader switch).

And it always bugs me when they pick Patriotism or Rich Land over Engineering Genius during round 1, and that happens a lot. They always pick wonders so that's not the explanation. Other Eng.Gen cards they also pick very highly.
 
 Thumb up
 tip
 Hide
  • [+] Dice rolls
Alexander Stumpp
Germany
flag msg tools
Flyboy Connor wrote:
Monte-Carlo Tree Search has no limited depth. It searches until the end of the game, using random moves.

While in theory the algorithm might very well be capable of that, I highly doubt that it really does this in the game.
Too many of the future possibilities are obscured by not knowing what cards there will be so I don't think there's much point to simulate everything to the end based on probabilites.

Knowing better about the evaluation-function would probably also help to improve my own play.
 
 Thumb up
 tip
 Hide
  • [+] Dice rolls
Alexander Stumpp
Germany
flag msg tools
In the meantime I have improved my play and moved up to hard difficulty aswell.

And I have to say the hard AI isn't that much better than the medium one. If anything it seems to play more predictable.

Medium AI would sometimes commit to seemingly a lot riskier plays and it actually would work out somehow.

Hard AI seems to have an agenda of never fall behind in military which lets it neglect culture for too long. So by being slightly ahead in culture production early on and then go along with their military-buildup, you force them to make even more military and stay ahead in culture.

They also seem to play fewer "impact of ..." events.

But they have some habits definitely worth copying. Their colony-bidding is much better than at lower levels and actually seems to make sense now and their scienceroduction:food:happyness-balance also seems a lot more solid.

But I've seen at least 2 occassion where they didn't spend all of their civil-actions and one occassion where they disbanded and then rebuild a unit in the same turn. Stuff like that really shouldn't happen.
 
 Thumb up
 tip
 Hide
  • [+] Dice rolls
Front Page | Welcome | Contact | Privacy Policy | Terms of Service | Advertise | Support BGG | Feeds RSS
Geekdo, BoardGameGeek, the Geekdo logo, and the BoardGameGeek logo are trademarks of BoardGameGeek, LLC.