Author: Michael Yee
Date: 14:57:10 04/27/05
Go up one level in this thread
On April 27, 2005 at 12:17:42, Rémi Coulom wrote: >On April 27, 2005 at 10:35:28, Michael Yee wrote: > >>Hi, >> >>I might have a chance to do some computer chess related research as part of my >>thesis for school. (Already planned are a couple topics related to "inferring >>human decision-making processes".) Computer chess will fit (in a loosely related >>way) if I look at techniques for learning parameters/strategies from a database >>of human GM games. >> >>Before I continue my literature search, etc., I just wondered what people with >>experience in this area think of the following ideas: >> >>- learn rules/functions for when to extend or reduce search along a given line > >You can take a look at this: >http://www.cs.ualberta.ca/~yngvi/Papers/isj03.pdf > This is a neat paper and seems similar to what I had in mind. But I suppose it could be extended even further by adding features like node counts for various branches and (previous) depths. I think the author of Hermann uses a neural network for time management and uses node counts for different branches as inputs. >>- learn when it's safe to prune a given line (related to previous idea) > >Probcut and multicut are based on ideas close to machine learning, if I remember >correctly. According to what I have read, Fabien Letouzey has also implemented >an original "history-pruning" scheme that seems efficient, but I do not remember >the details. > I guess parameters for these type of schemes could be estimated from GM databases. >Also, I believe it might be a good idea to use machine-learning techniques to >learn extended-futility-pruning heuristics. > >>- learn parameters for static (leaf node) evaluation function (although a lot >>has already been done here, I think) > >Yes, this has already been done. For chess, check Baxter & Tridgell, and Beal & >Smith. The results were not impressive. > Okay, so more work could be done here. I'm considering using very large (hopefully representative) training sets, not just self-play (as long as I can find the computing time!). >>- learn/construct/discover features for a static evaluation function (some >>not-so successful work may have been done here with neural networks?) > >I believe that finding good features necessarily requires the intervention of a >human expert. But it might be a good idea to try to find ways to help a human >expert find out what kind of knowledge is missing most in the evaluation >function. I have no idea how, though. > >>- learn rules for move ordering (i.e., that try to search best moves from a >>given node first to achieve more efficient cut-offs) > >The history heuristic and its variations are in the spirit of this. > >> >>Specifically, I'm curious which areas are already "mature", which seem >>promising/new, or even if you have any other ideas/references. > >Computer chess as a whole is extremely mature, so it is hard to make significant >contributions. > >During the Ramat-Gan World Championship, I asked Amir Ban and Shay Bushinsky >wheter they were using machine learning techniques in Junior. They answered >something like "yes, but not the machine-learning techniques you know". That >might be an interesting indication that there is potential for investigation in >this direction. > Their comment is very intriguing... Of course, their ML techniques happen to be coupled with a super strong pseudo-conventional engine (I'm guessing). >> >>Thanks! >> >>Michael > >Good luck with your research, > >Rémi Thank you for your comments and ideas. I'll keep the board posted on any progress. Michael
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.