Why is a classifier better than having a lookup table indexed by OurLastMove, OppLastMove, ProposedNextMove that returns the Wins / Trials experienced when ProposedNextMove is played after the sequence OurLastMove, OppLastMove?
Are the training cases for your classifier selected from only the UCT nodes, or also from playout nodes? Is the output of your classifier used to initialize the Wins / Trials values for legal moves in new UCT nodes? Is that done by assuming a fixed number of trials (how many?) and setting Wins = ClassifierOutput * Trials? Is that the only use of the classifier in the system? From: [email protected] [mailto:[email protected]] On Behalf Of Peter Drake Sent: Wednesday, June 29, 2011 11:20 AM To: [email protected] Subject: Re: [Computer-go] A Linear Classifier Outperforms UCT on 9x9 Go On Jun 28, 2011, at 9:39 PM, Imran Hendley wrote: Hi, long-time lurker and occasional poster here, Thank you for the paper. I hope you don't mind me asking a few very basic questions, since I am having trouble understanding exactly what you are doing. Let's say we are using a linear classifier. Then our output (the predicted move) should look like: argmax_i (y[i]), where y[i] = w1[i] . m1 + w2[i] . m2 + b Where each w[i] is a weight vector for location i on the board, the m's are the (column) input vectors (which I assume are 1 at the move location and zero elsewhere), and b is the bias term. There is a separate bias for each move, so b in your formula should be b[i]. To train our classifier online, we want to do something like: (1) Generate a prediction for a training example. (2) Calculate the error. (3) Update the feature weights. (4) Repeat. If I understand, online training happens during the course of one game, as we are playing. Moreover, we are using our classifier to generate moves to select in the first phase of our simulation, as a replacement for MCTS, and before playouts. Correct. Now this is where I have to start guessing the details. Are our training examples playouts, and is our error function just 0 if the playout wins, and 1 if it loses? The "correct output" is 1 if the playout wins, 0 if it loses. The error is the difference between the correct output and the actual output. And as we run more playouts, the classifier will update its weights and select a different sequence of moves in the first phase of our simulation (analogous to selecting different paths down the search tree based on node scores in MCTS)? And when we use up our allotted time for one turn we just return the next move (from the current position) that our classifier predicts, based on its current weights? We tried this, but the classifier fluctuates quite a bit. (This is, we think, a desirably property to keep up exploration.) Instead, we choose as the actual move the move through which the most playouts were played.) The paper says we fix the number of moves we select with the classifier before running playouts (unlike starting from the root and expanding in MCTS). This is where things start getting really fuzzy for me. Do we propagate the results of a playout back up this sequence? i.e. if we get a win, do we perform updates of our classifier for each two-move sequence in the full sequence? Yes. The classifier therefore learns from the entire playout, not just from moves generated by the playout. (This is vaguely analogous to RAVE.) I would really like to get to the deeper questions about interpreting what is really going on, but I first need to make sure I am on the right page here. Sincere apologies for the stupid questions. I really hope my understanding didn't get derailed so early on that most of my questions in this message are gibberish. But I did want to show that I actually made a concerted effort to understand the paper before asking what on earth it is all about! No problem -- we look forward to any insights you can offer! Peter Drake http://www.lclark.edu/~drake/
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
