Timeline for A purely self-trained chess AI
Current License: CC BY-SA 3.0
9 events
when toggle format | what | by | license | comment | |
---|---|---|---|---|---|
Jul 8, 2012 at 15:08 | comment | added | thb | I have updated the AI question in light of your answer. The update is not short so, at your option, when you have some time, you can review it to the extent to which it interests you. | |
Jul 7, 2012 at 15:27 | comment | added | Greg E. | @thb, I'm no expert, but I think that's the case. Even the AI you linked to is constrained by programmer biases. If you read the original research paper, it notes that the "feature vector" training the evaluation function comprised board features that "were carefully designed by hand." That is, the programmer still has to specify the set of static positional factors the AI bases its decision-making on. The major advantage of a neural net for this particular project, I think, is that the training can be parallelized, allowing for asynchronous processing of massive amounts of games efficiently. | |
Jul 7, 2012 at 15:08 | comment | added | Eve Freeman | It might be interesting to see what a genetic minimax engine comes up with, for example, if you were to start all pieces out with the same material value, based on win/loss/draw, and let the material values mutate. I'm sure engine creators have already tried tweaking these, like making knights 2.9 and bishops 3.1 pawns. | |
Jul 7, 2012 at 15:02 | comment | added | thb | @GregE.: No plausible chess AI? I suspect that you are right; for, if you were wrong, then such an AI would have been made by now. | |
Jul 7, 2012 at 15:00 | comment | added | thb | @WesFreeman: That's right. It would develop its own evaluation function. Note that this is not my own idea, but appears to be what the pure chess AI linked in the original question tried and failed to do. | |
Jul 7, 2012 at 14:01 | comment | added | Greg E. | @thb, as a programmer, I think the issue with your notion is that, as far as I can see, no plausible chess AI can begin with a totally blank slate for an evaluation function. One could write an AI that analyzes games for patterns and uses statistical/probabilistic methods (e.g., Bayesian inference) to fine-tune its valuations and decision-making, but the programmer has to identify what motifs, positional factors, move sequences constitute said patterns and by what criteria to assess them. In other words, the basic core of the evaluation function would still need to be human-designed. | |
Jul 7, 2012 at 13:47 | comment | added | Eve Freeman | So you're saying it would develop its own evaluation function? | |
Jul 7, 2012 at 12:41 | comment | added | thb | The answer is appreciated. What I believe that I had in mind was an AI that (a) possessed a minimax capability but (b) lacked a predetermined evaluation function. Such an AI would necessarily solve so small a game as tic-tac-toe by pure minimax. In chess, a game only theoretically susceptible to minimax, the AI would evaluate not the present position on the board but future positions, after which minimax would choose the move. It might loosely be said that Nimzowitsch revolutionized chess by spurning known evaluation heuristics. If so, then could a machine do likewise? | |
Jul 7, 2012 at 5:16 | history | answered | Eve Freeman | CC BY-SA 3.0 |