Approaches

N-tuple leaf

Does a learned evaluation, with the search unchanged, beat the hand-written leaf, and does temporal-difference self-play learn anything at all through this pipeline?

The mechanism

Tuples of squares over the observation index weight tables; the value is their sum. Tables are trained by TD(λ) from self-play games of an ε-greedy afterstate player, using the environment bindings. The concept page has the alphabet and tuple placement.

The plan

Train for a fixed budget of games, evaluate the n-tuple afterstate player against greedy every 100,000 games, and stop when the curve flattens for three checkpoints. Then register the leaf swap: the same expectimax search with the learned leaf against the same search with the hand-written leaf, 200 paired games.

Decision rule

The learned leaf is promoted if the paired contrast against the hand-written leaf under identical search has a 95% interval above zero, and the learned afterstate player alone beats greedy with an interval above zero. If the second holds and the first does not, the tables are kept as an evaluator for the RL rung's comparisons but not as the search leaf.

What could disprove it

TD self-play from random setups may learn to value stalling. A learned leaf that raises the draw share against greedy without raising the win share is recorded as a failure of the objective, not of the tables.