Bayesian POMDP

The performance gap between the clock-state agent and the Bayesian agent is real but modest, and grows most in sparse environments where history-dependent adaptation confers the largest marginal advantage. The Bayesian POMDP baseline neglects temporal correlations between successive odor observations, which are present in real turbulent flow; incorporating them would improve accuracy but substantially increase computational cost. The Bayesian POMDP agent maintains a posterior distribution over source locations, updated by Bayes' rule after each observation using empirically estimated detection likelihoods. The Bayesian agent produces broad recovery trajectory distributions in which all five geometric metrics vary widely with prior detection history, even when the triggering condition is identical.