About Chase

Chase plays Othello with three convolutional networks, a PUCT search over them, an opening book, and an exact endgame solver. The same engine runs here in your browser as the one measured in the results below — the search visits the same nodes, and the solver returns the same scores.

How a move is chosen

Up to twelve discs the move comes from an opening book, because the networks were trained on positions from twelve discs upwards and have never seen an opening. With sixteen or fewer empty squares the whole move is solved exactly, so the number you see is the final disc difference under best play, not an estimate. Everything between the two is searched.

What the numbers mean

During search each square shows the win rate for the side to move and how many simulations went into it. Shading is relative to the strongest move on offer, so two moves that are close look close. In the endgame the numbers switch to exact disc differences, and in the opening to the book's own evaluation.

Where it came from

Training data came from an engine, from WTHOR game records, and from positions labelled by hand. The current networks were trained on about forty-eight million positions per stage. Doubling the data was worth roughly five points of win rate head to head; raising the search from four hundred to twenty-five hundred simulations was worth twenty-one.

This section is a placeholder — the real write-up goes here.