Topic: News about AI
Tibor Topic Opener |
Posted at: 2017-08-12, 18:05
WOW, I believed you gave up after second round. Are you able to continue with the training? I have only few maps used for training so manual training provides variability here and is huge benefit.... Top Quote |
Nordfriese |
Posted at: 2017-08-12, 18:42
I´ll have lots of spare time for the next few weeks, so I´ll continue with the training. I enjoy it – one AI player has already shown me a way how to get level 8 empire soldiers faster than I ever could Top Quote |
Tibor Topic Opener |
Posted at: 2017-08-12, 19:34
Oh, this is great. I would be relieved. There are couple of issued that should be addressed. One of biggest is the variation. It is not good that generated AI players are too diverse. Players expect roughly equivalent performance. The main way how to achieve that is through mutation rate as shown also in logs:
So probably version in trunk should have probability 1/200 (200 is right now maximum, but can be changed) and training mode as by now in range 1/125-1/175. Because in training we need diversity.... What is your opinion about this aspect? Mainly diversity / variability problem... Top Quote |
Nordfriese |
Posted at: 2017-08-13, 09:22
I agree with you. If training has a small diversity, then the chances that a slightly better AI turns up soon are high but the difference is not that great. If the mutation rate is very high, it may take longer until a really good AI appears but it will be significantly better than its parents. It´s more of a question whether one wants many short or few long training rounds, and that´s a question of preference of the trainer, in my opinion. I´m satisfied with the current mutation rate. In normal games, this is a different matter, of course. Higher variability means a higher risk of having a weak AI, and also a higher risk of having one weak and one strong AI, which is something I wouldn´t appreciate. I found this recently in a game I played as frisians – after I had defeated the ridiculously weak barbarians I was beaten by a surprisingly strong empire AI. A smaller variability of 1/200 or less would be better here. Top Quote |
GunChleoc |
Posted at: 2017-08-13, 10:11
You cheat! Thanks for helping with the training, and thank to Tibor for improving our AI Busy indexing nil values Top Quote |
hessenfarmer |
Posted at: 2017-08-13, 14:39
@ Nordfriese: Top Quote |
Ex-Member |
Posted at: 2017-08-13, 15:58
I have been having problems getting widelands to run at all following some library updates, solved now and I just lost my first game ever which shows a great improvement in AI in general, congratulations, it can only get better (harder) from here on. I want to carry on training but am being particularly dense it seems, I am running 8419 and games start and run OK, on first run a folder ~/.widelands/ai is created but no files are added, Nordfriese imlies they are for him but I am not sure how I can generate my own, it seems that traingingmode must be turned on but I see no instructions about how to do it, Tibor implies it has yet to be implimented. Top Quote |
GunChleoc |
Posted at: 2017-08-13, 16:54
A the moment, training mode can only be switched on by changing a variable in the source code and recompiling. A command line switch is planned, but hasn't been implemented yet. Busy indexing nil values Top Quote |
Tibor Topic Opener |
Posted at: 2017-08-13, 18:06
branch with ai training switch is already functional, but needs some polish. I will propose its merging soon. @ hessenfarmer: If you have a specific map that is difficult for AI you can approach Nordfriese and ask him to train AI on this special map... Top Quote |
Nordfriese |
Posted at: 2017-08-13, 18:50
@hessenfarmer: I guess it will take much more training until the AI works well on your map. Barbarian AI is currently the least deadlock-prone, but it finds it hardest to establish a good economy and is almost incapable of training its soldiers. My training currently focuses on improving barbarians in these two respects. If you give me a link to this map, I´ll use it in my training rounds until AI performs reasonably well on it. Top Quote |