Policy Atlas
What kinds of policies did the solver learn? 6,000 real information sets from the 2M-iteration release strategy, projected to 2D (deterministic PCA over their action distributions). Each point is an actual stored infoset.
f (fold)c (check/call)x… / b… (raise/bet)a (all-in)
Distance here is similarity of ACTION DISTRIBUTIONS under a fixed action basis — not poker-state similarity. The tight clusters are near-pure policies (always-fold, always-call); the connecting cloud is genuine mixing.
How mixed is the learned strategy?
| population | near-pure (≥95%) | moderately mixed | highly mixed (<70%) | mean entropy |
|---|---|---|---|---|
| all infosets | 36.5% | 26.5% | 37.0% | 0.76 bits |
| preflop | 34.2% | 26.0% | 39.8% | 0.84 bits |
| flop | 36.0% | 25.1% | 38.9% | 0.78 bits |
| turn | 35.7% | 27.2% | 37.1% | 0.76 bits |
| river | 36.9% | 26.6% | 36.5% | 0.75 bits |
stored abstract infosets, not encountered-hand frequency.