Evaluating agents

Evaluating agents

src/eval.rs answers the practical question the analytics tools don't: if two agents actually played, who would win, and by how much? All three functions panic if profiles.len() != num_players() (and play_match also requires game.num_players() == start.num_players()), if the game is malformed (sample_chance / step fails), or if a subgame fails to terminate within DEFAULT_MAX_STEPS (10,000,000). The try_* variants return Result instead of panicking.

• play_out(game, profiles, rng): sample one trajectory of a single subgame under the per-player policies; returns the terminal payoffs. Stochastic. Repeat and average for a win rate. Panics on a player-count mismatch.
• play_match(game, profiles, start, transition, rng): chain subgames via a caller-supplied transition and play the full match (many rounds), accumulating terminal returns. transition(&terminal, &actions) -> Option<G> encodes how one resolved subgame advances to the next (None ends the match). Panics on a player-count mismatch, and also requires game.num_players() == start.num_players().
• head_to_head(a, b): the exact expected payoff of profile a (player 0) versus b (player 1) via full-tree expectation over the joint profile, with no sampling noise. For non-enumerable-chance games it falls back to a fixed-seed 64-sample Monte-Carlo walk. Returns (payoff_to_0, payoff_to_1). Panics on a player-count mismatch.

A single play_out is a strict subset. It is equivalent to play_match with transition returning None immediately.

Fallible variants

• try_play_out / try_play_match / try_head_to_head return Result instead of panicking. Use these when you must not panic, including when the player count is not known to be correct ahead of time.