Sublinear regret for learning POMDPs

We study the model-based undiscounted reinforcement learning for partially observable Markov decision processes (POMDPs). The oracle we consider is optimal policy of POMDP with a known environment in terms average reward over an infinite horizon. propose algorithm this problem, building on spectral method-of-moments estimations hidden models, belief error control POMDPs and upper confidence bound methods online learning. establish regret O ( T 2 / 3 log ) $O(T^{2/3}\sqrt {\log T})$ proposed where This is, to best our knowledge, first achieving sublinear respect general POMDPs.