Latest Posts

Topic: AI-Training

Tibor

Joined: 2009-03-23, 23:24
Posts: 1377
Ranking
One Elder of Players
Location: Slovakia
Posted at: 2019-04-26, 14:59

WorldSavior wrote:

Okay, the problem here is that the best AI could be "doubled" by a pair of other AIs, so it loses.

You are right and considering big randomness, there is no guarantee that best performer is indeed absolutely best. F.e. if all your neighbors are week, you might look like extra good player/AI. But this is how it works also in nature...

I'd rather pick the four best performers in total than the best performer from each start spot.

BUT - if you test on "normal" maps, not all starting posts are equal and you would automatically prefer AI players from good position(s)

But on periodic map this approach would be completely fine....


Top Quote
WorldSavior
Avatar
Joined: 2016-10-15, 04:10
Posts: 2091
OS: Linux
Version: Recent tournament version
Ranking
One Elder of Players
Location: Germany
Posted at: 2019-04-26, 15:11

Tibor wrote:

WorldSavior wrote:

Okay, the problem here is that the best AI could be "doubled" by a pair of other AIs, so it loses.

You are right and considering big randomness, there is no guarantee that best performer is indeed absolutely best. F.e. if all your neighbors are week, you might look like extra good player/AI. But this is how it works also in nature...

It works like that in nature, but for our AI training it's bad.

I'd rather pick the four best performers in total than the best performer from each start spot.

BUT - if you test on "normal" maps, not all starting posts are equal and you would automatically prefer AI players from good position(s)

But on periodic map this approach would be completely fine....

Did you find out that the starting positions of Desert Tournaments have big quality differences? For me it looks like they are very similar.

Edited: 2019-04-26, 15:20

Wanted to save the world, then I got widetracked

Top Quote
Tibor

Joined: 2009-03-23, 23:24
Posts: 1377
Ranking
One Elder of Players
Location: Slovakia
Posted at: 2019-04-26, 15:25

WorldSavior wrote:

It works like that in nature, but for our AI training it's bad.

AI has like 4 more less independent lines/variants. Sometimes a variant dies-out, and another split into two/three lines.

I hope it is clear the AI regardless of starting position pick random one from 4 WAI files...

So inefficiencies does not kill entire training.....

Did you find out that the starting positions of Desert Tournaments have big quality differences? For me it looks like they are very similar.

Yes, some normal maps have minimal differences between starting points. But why to bother to split maps into "equal starting points" and "not equal starting points"?

Edited: 2019-04-26, 15:26

Top Quote
king_of_nowhere
Avatar
Joined: 2014-09-15, 18:35
Posts: 1668
Ranking
One Elder of Players
Posted at: 2019-04-26, 15:40

I agree that desert tournament is a great map for training, perhaps more variety of maps would be welcome. Unfortunately, one thing that is really difficult to do with the AI is that you have to adapt your strategy to the map, and the AI can't really do that.

Desert tournament is great in that it balances early game with mid-late game, but it does llittle for a huge map with really late game big armies. I think ice wars would be a good map to temper the training with some adaptation for large maps.

WorldSavior wrote:

It works like that in nature, but for our AI training it's bad.

the idea is that an inefficient training cycle that can be fully automathized and run authomatically a huge number of times is better than a more refinined process that requires human input and is slower. Inefficiencies caused by statistics tend to even out in the long run. It's the zerg rush equivalent of problem solving.

Did you find out that the starting positions of Desert Tournaments have big quality differences? For me it looks like they are very similar.

there are some small differences, one of the players have slightly more land and one slightly less. Not sure oon specifics, I remember during the 2015 tournament I was able to make one extra farm with a couple of starting places. Difference is minimal though, can be considered equal for all practical purposes. If you can train the AI to the point where that tiny difference actually matters, it's already a great success.


Top Quote
teppo

Joined: 2012-01-30, 09:42
Posts: 423
Ranking
Tribe Member
Posted at: 2019-04-27, 15:49

Tibor wrote:

The training should look like - you pick a map with 4 starting positions and run it like 15 times, so you will have 15 wai files from position 1, another 15 from position 2 and so on. Then from each 15 pick the best performing one and place it as a new source WAI file...

Who does the scoring? I write down a verbal description of the occupied slot goodness of victory each time, and compare the essays to find the best ones after running all rounds?

Of course there is not link between player starting post and source wai files.... Also keep in mind that generated wai is usually worse than one in source wai files, this is how genetic mutations work, only sometimes it can be better... this is why you should pick the best one from bigger group of generated WAIs....

Is there any way to check whether the new genes are better than old ones?

Who makes the bundled WAI files? Are they good-enough for build 20?


Top Quote
stonerl
Avatar
Topic Opener
Joined: 2018-07-30, 00:03
Posts: 327
Ranking
Tribe Member
Posted at: 2019-04-27, 17:58

As far as I understand, the major problem is that the AI would have to be trained on a map that covers all hurdles it could run into. e.g. if I train only on "The Crater", the AI would perform exceptional on this map but not on others?

Some kind of automation would be nice. I can spend a lot of CPU time, but my personal time is quite limited. We then could train the AI for specific maps and load the WAI-files for the trained map. If no map specific WAI-files a present we then would fall back to standard WAI-files. Would such a thing possible?


Top Quote
Nordfriese
Avatar
Joined: 2017-01-17, 18:07
Posts: 1929
OS: Debian Testing
Version: Latest master
Ranking
One Elder of Players
Location: 0x55555d3a34c0
Posted at: 2019-04-27, 18:16

the AI would have to be trained on a map that covers all hurdles it could run into. e.g. if I train only on "The Crater", the AI would perform exceptional on this map but not on others

In my opinion, a map with many challenges is poorly suited. E.g. if one AI increases its skill at seafaring but loses skill at coping with little space, and the other AI is the other way around, which one is better?

In my opinion, it makes more sense to use many different maps each one of which has one big challenge, and the winning AIs are selected by how well they cope with the map´s particular challenge. (Regarding the other aspects in which the AI has to be trained, the map would have to match the AI´s current skill level though, so it doesn´t lose a skill again if it doesn´t need it there.)


Top Quote
WorldSavior
Avatar
Joined: 2016-10-15, 04:10
Posts: 2091
OS: Linux
Version: Recent tournament version
Ranking
One Elder of Players
Location: Germany
Posted at: 2019-04-27, 18:17

Tibor wrote:

WorldSavior wrote:

Did you find out that the starting positions of Desert Tournaments have big quality differences? For me it looks like they are very similar.

Yes, some normal maps have minimal differences between starting points. But why to bother to split maps into "equal starting points" and "not equal starting points"?

For example in multiplayer matches it is useful to have equal starting points. For AI-training it should be useful as well.

king_of_nowhere wrote:

Desert tournament is great in that it balances early game with mid-late game, but it does llittle for a huge map with really late game big armies. I think ice wars would be a good map to temper the training with some adaptation for large maps.

At the other hand, distances between players on ice wars are not huge, so it can also be considered as mid-game map.

stonerl wrote:

As far as I understand, the major problem is that the AI would have to be trained on a map that covers all hurdles it could run into. e.g. if I train only on "The Crater", the AI would perform exceptional on this map but not on others?

I think so. At the other hand it is not certain that a AI which is trained on a very hard map can play well on an easy map, is it?

By the way, is anyone interested in creating a map which is very good for AI training?


Wanted to save the world, then I got widetracked

Top Quote
teppo

Joined: 2012-01-30, 09:42
Posts: 423
Ranking
Tribe Member
Posted at: 2019-04-27, 18:41

WorldSavior wrote:

As far as I understand, the major problem is that the AI would have to be trained on a map that covers all hurdles it could run into. e.g. if I train only on "The Crater", the AI would perform exceptional on this map but not on others?

I think so. At the other hand it is not certain that a AI which is trained on a very hard map can play well on an easy map, is it?

On the positive side: If somebody would go though the huge amount of work to make the training to happen by itself (=given enough CPU cycles, evolution progresses itself), then it would be possible for (some) maps to contain .wai files optimized for that map.

Edited: 2019-04-27, 18:45

Top Quote
Nordfriese
Avatar
Joined: 2017-01-17, 18:07
Posts: 1929
OS: Debian Testing
Version: Latest master
Ranking
One Elder of Players
Location: 0x55555d3a34c0
Posted at: 2019-04-27, 18:47

On the other hand, the default wai files will be improved frequently, but the map´s wai replacement files probably won´t be. At some time the default AI might then perform better than the map´s. I´m against map-specific AI files as it would require too much maintenance to be efficient…


Top Quote