Baseball, An Extensive-Form Game-Theoretic Duel
arXiv:2607.29041
2026
Training
2 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
The paper provides an exact sequence-form treatment of finite imperfect-information zero-sum games, together with a count-state backward recursion and action-specific one-step deviation values. The transferable asset is not baseball itself, but the combination of realization-plan flow constraints, saddle-point continuation values, and nonnegative deviation gaps that quantify how exploitable each local decision is. These constructions can be used to train paired neural policies or hierarchical routers with explicitly valid probability flows and to prioritize optimization on locally exploitable actions. The ideas are most relevant to adversarial training, multi-agent policy learning, and tree-structured or hierarchical neural decision modules.
Ideas from this paper
Unverified
2026
Add an action-level exploitability penalty to alternating training of two neural policies that play against each other. For each observed state, estimate the value of forcing every available action against the opponent's current policy, then penalize positive gaps from the player's minimax value rather than relying only on the sampled action or episode return. This should expose locally exploitable decisions earlier and reduce oscillation between adversarial policies.
Useful6/10
Difficulty5/10
Novelty5/10
Unverified
2026
Parameterize a tree-structured policy through realization weights satisfying sequence-form flow conservation, instead of independently predicting probabilities at every node. Conditional action probabilities are recovered by dividing a child sequence weight by its parent weight, guaranteeing globally consistent probabilities and avoiding invalid or contradictory branch masses. This is suitable for hierarchical RL policies, adaptive computation trees, and neural routers with sequential gating…
Useful5/10
Difficulty4/10
Novelty6/10