Tuesday, December 03, 2002
( 4:53 PM ) Matt
A while ago I talked about how there are a lot of different ways to teach neural nets, or similar AIs, to play Go. Studying masters' games and playing competitively are two common ones. Today I'm going to talk about a couple more strategies.
It's been shown that teaching a neural net by having it play against really good opponents is not a good way to teach it. It needs to start with easier opponents. It's also been shown that if you teach a neural net by showing it a whole lot of masters' games, then it will play go that looks like professional quality go, except for a few glaring, fundamental, errors. Perhaps what it's missing is learning from games which are closer to its level of playing. So, instead of teaching the net only from masters' games, perhaps it would be better to teach it from games between 20k players, and then slowly work up to professional games.
Group study has proven successful among real players. How can we adapt this to artificial players? Here's my proposal. Take a group of neural net players, and have them vote on moves, or otherwise agree collectively on what the best next play is. Play out an entire game this way and find the winner. Then, go back along the tree until you find a choice made by the losing player. (i.e. a move suggested for the loser which was not the one the group collectively chose.) Then play out the game from that point. Repeat this process with all the choices of the losing player. If a different choice leads the formerly losing player to win, then learn from this.
The second learning technique here is what I hope people will do with the game exploring which I proposed earlier. Maybe AIs can join in on this too and learn along with the people. Again, it's probably best to study with a group of players of roughly equal rank.
One difficulty in the group study proposed above is that it requires the game to be played to the end before it can be determined if the move is good or bad. For working out the quality of joseki (beginning game) moves, this could be very tedious. It seems that it is also important for a player to be able to estimate the score before the game is finished. ... More on this later...# -
Comments: Post a Comment