Publications
-
Adaptive multi-agent driving with population play.
We train ego policies against a learned, conditioning-aware co-player population in a multi-agent driving simulator and study held-out adaptive-adversary generalization. The work characterizes how architecture, co-player diversity, and reward shaping interact to determine whether a policy adapts within a trial, and proposes a decisive replication protocol that disentangles within-config variance from cross-cell effects.
-
Finite-size scaling of Schelling-type segregation.
A finite-size scaling study of the Schelling segregation model and its multi-radius generalizations. We extract critical exponents at four interaction radii on lattices up to L = 2048, test universality across radii, and verify hyperscaling. The collapse establishes a noise floor for the segregation order parameter and clarifies which prior claims survive at sufficient n.
-
Adaptive sensing driven by prediction error.
We design adaptive sensing strategies that exploit a learned forward model: each next measurement is chosen to maximize the expected reduction in predictive uncertainty, concentrating sensing on regions where the world's structure has not yet been captured. The policy reaches calibrated coverage with a fraction of the queries needed by uniform or active-learning baselines, and the analysis links the sample budget to the intrinsic dimension of the underlying signal.