Bot training

Loading…

Settings for the next run

Learning rate Time limit, minutes (0 = until I stop) Test every (minutes) Test every game type every (minutes, 0 = never) Game types (4a: 2 v 2 two clean, 4b: clean and dirty, 2a/2b: 1 v 1) Lookahead while training (deals, moves, replies; empty = none) Share of decisions that look ahead Train with the game's charges (guard, discipline, risk) Start again by itself if it stops for anything but the Stop button

This run against the network before it

Points per round ahead of the network it started from, at each test. The bars show how far luck can move a test; the line is the average of the last five. This is the number that says whether the run is paying.

Each testAverage of the last five

How it plays

From the newest test.

Strength against SimpleBot

Points per round ahead of SimpleBot, over all training time. Higher is better.

End-of-round test (2,000 deals, ±6)Mid-run test (300–400 deals, ±16)

Gain over the network before it

Points per round each round gained against the network it was compared with. Below zero means it got worse.

Rounds played per second

Self-play speed, every 15 seconds while a round runs.

GPU temperature, °C

Shown for information; training does not pause for heat.

Learning error

How far the network's guesses are from real results. It flattens early; strength is the real measure.

Settings of the run

All rounds

Every training round so far, newest last. The last four columns are the network's own figures in its final test against SimpleBot.