Combinatorial Causal Bandits

In combinatorial causal bandits (CCB), the learning agent chooses at most K variables in each round to intervene, collects feedback from observed variables, with goal of minimizing expected regret on target variable Y. We study under context binary generalized linear models (BGLMs) a succinct parametric representation models. present algorithm BGLM-OFU for Markovian BGLMs (i.e., no hidden variables) based maximum likelihood estimation method and give analysis it. For special case we apply inference techniques such as do calculus convert original model into model, then show that our another regression both solve variables. Our novelty includes (a) considering intervention action space general graph structures including ones (b) integrating adapting diverse studies online influence maximization, (c) avoiding unrealistic assumptions (such knowing joint distribution parents Y all interventions) factors exponential size prior studies.