Opponent adaptation

How quickly should a player change its strategy when an opponent behaves differently?

The decision at the table

The same refused offer means different things from a conservative builder, a reciprocal trader, or a player that ignores chat. A flexible player should learn those differences without turning one observation into a personality diagnosis.

The mechanism to test

Use a small set of behavioral hypotheses and update their weights from observed decisions. Keep uncertainty and a fallback response. Distinguish factual adaptation from accommodating pressure: better predictions may justify changing a move even when the player’s objective remains exactly the same.

Proposed experiment

Compare a fixed opponent model, an online model with bounded updates, and a model with change detection when behavior shifts. Hypothesis: online adaptation improves predictions and wins against unfamiliar opponents while avoiding large losses to misleading early behavior.

Plot prediction error through each game and cross-play outcomes by opponent class. Track how often a model changes its preferred action. Evaluate on withheld policy versions and messaging styles, not only alternate prompts from the same template.

First ablation

Modelling opponents as leaf-guided greedy players instead of the fixed builder changed the contrast by -0.008 wins per game (95% interval -0.111 to +0.096) at roughly double the decision time; against actual builders the builder model is the better-calibrated one. Details and every run ID are in the ablation report; this is a development-tier result on one lineup.

Reading the next build

Predicting each opponent's next build from public information gave one term a development signal that fresh seeds erased. A denial leaf term discounting sites opponents are heading for read +0.061 on seeds 0 to 63 and -0.025 on seeds 64 to 127; the predicted opponent model and the predicted space threat never separated from zero. The predictor named 70% of the fast builder's next builds and 28% of another search's, so it reads the weakest seats best. The report also measures a slot-order bonus of +0.172 wins per game for the seat acting immediately before its twin on the development seeds, which any single-order contrast there carries.

Third round: prediction under the learned leaf

Retested with the learned tables as the leaf, every build-prediction switch stayed inside its interval. The predicted space threat read +0.047 wins per game (95% interval -0.004 to +0.098) on a population lineup of three searches, +0.014 on the development boards, and -0.006 on fresh builders boards; the predicted opponent model read -0.010, +0.000, and -0.004; and the denial term, measured on the quarter blend of tables and hand-written terms, fell to -0.037 (-0.071 to -0.003) on fresh seeds at about double the decision cost, after its hand-leaf development gain of +0.061 had already failed fresh seeds. The predictor names a tables search's next build in 20.0 percent of turns, worse than the 28.5 percent it scored against the hand-leaf search, so the seats that win these tables are even harder to read than before. The report has the pairs, the hand-leaf comparison, and the browser reading: no switch is worth enabling.

What could disprove it

Adaptation can become exploitable if a rival intentionally teaches the wrong pattern. Report worst-opponent performance. Private identity attributes are unnecessary: game behavior and the explicitly available memory are the research inputs.

All approaches · Player’s guide · Experiment program

Tracked investigations