Maps the engine's evaluation or MCTS root values to a [0, 1] estimate
for player. The interpretation depends on the selected method.
Usage
mxo_win_prob(
game,
player = mxo_to_move(game),
method = c("mcts", "heuristic"),
...
)Arguments
- game
An
mxo_gameobject.- player
Integer scalar,
1L(X) or2L(O). Defaults to the player to move. Integer-valued numeric scalars are accepted; missing, fractional and other values are rejected, including for terminal games.- method
One of
"mcts"or"heuristic". Default"mcts".- ...
Additional arguments forwarded to
mxo_mcts()whenmethod = "mcts"or tomxo_evaluate()whenmethod = "heuristic".
Methods and calibration
The default method = "mcts" converts visit-weighted rollout values to
an expected score, counting a draw as half a win, and applies smoothing
towards 0.5. Unfinished rollouts use a bounded heuristic evaluation.
This is not the probability of a decisive win alone.
method = "heuristic" uses the fitted logistic calibrator shipped in
R/sysdata.rda. Its build recipe, data-raw/make_calibrator.R, uses a
small batch of random-policy self-play on a 3^3 board. Calibration fits
exclude draws, so this estimate concerns wins among decisive outcomes in
that training setting. It is not validated across all board configurations
or against human players. Legacy logistic constants are used only if the
internal fitted calibrator is absent.
Both methods return 1 for a terminal win, 0 for a terminal loss and
0.5 for a terminal draw, without running a search or calibration fit.
Examples
g <- mxo_new_game(n = 3L, k = 3L)
mxo_win_prob(g, method = "heuristic")
#> [1] 0.542849