One iteration (self-play, training, evaluation) counts as one generation. Each iteration the Monte Carlo search, guided by the value net, plays self-play rounds, the net is retrained on the recent games, and the new candidate net must beat the kept net on fresh paired deals (the gate) to replace it. Every iteration the kept net's search is also measured against the base algorithm (the game's medium bot) and, where reported, hard, on fresh paired deals like the search bots; the numbers are shown as an approx win %, (place score + 1) / 2, so 50% = level with the opponent; the advantage is the net's win % minus the opponent's, in percentage points, with its 95% margin. The kept net is the best net so far; the panel says since which iteration it holds, its shape, its latest scores, and by how much it beat the previous best when it was promoted.
3 runs = 3 independent nets; each keeps one best net, replaced when a better one wins.
Running now
Kept net (the best so far), per run
Approx. win % of the kept net against the base algorithm (50% = level), at each evaluation
Is the kept net beating the base bot? Above the dashed 50% line = yes, and the higher the better. Each dot: the kept net plays fresh paired deals against the base bot (the same cards dealt twice, seats swapped); win % = (score + 1) / 2.
The kept net is measured every iteration: each big dot is one evaluation with its 95% margin as the vertical bar, and a straight line joins one evaluation to the next. The x axis is shared with the gate chart, from iteration 0 to the latest. Hover for the values.
The gate: each iteration's candidate net against the kept net (50% = level; above 50% it is promoted)
Each iteration we train a new candidate net from recent self-play. It plays the current best (kept) net; if it wins, it becomes the new kept net (white ring).
One point per iteration: the candidate's approx win % against the net kept so far on fresh paired deals, the thin bar is the 95% margin (click a legend name to show that run alone, and the margin is shaded). A white ring marks a promotion.
Approx. win % of the kept net against hard (50% = level), at each evaluation
Is the kept net beating the hard bot, the strongest one? Above 50% = yes.
Seconds per iteration
How long one iteration takes. Flat = steady; it only matters for planning.
Training loss per iteration
How well the net fits what it has seen. Lower = a better fit; it does not by itself mean a stronger player.
Scores every iteration
Approx. win % of the kept net against the base algorithm (and hard), measured at the end of that iteration (– where none was made). Green: the advantage (in the hover: the net's score minus its opponents' average, which is not the same as the win % shown, since there are several seats) clears its 95% margin above zero; red: it clears it below zero; grey: within the margin (no real difference). ★ bold: the run's best so far; ⬆: that iteration's candidate was promoted. Click a column header to sort by it, again to reverse. Raw place scores in the hover.