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 random deals, each played twice with seats swapped (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 random deals like the search bots (each deal is played twice with seats swapped: once with the net in a seat, once with the other bot in that same seat); 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.
Approx. win % of the kept net against medium (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.