Towards Generalization and Efficiency in Reinforcement Learning

Abstract

Different from classic Supervised Learning, Reinforcement Learning (RL), is fundamentally interactive : an autonomous agent must learn howto behave in an unknown, uncertain,and possibly hostile environment, by actively interacting with the environment to collect useful feedback to improve its sequential decision making ability. The RL agent will alsointervene in the environment: the agent makes decisions which in turn affects further evolution of the environment.Because of its generality– most machine learning problems can be viewed as special cases– RL is hard. As there is no direct supervision, one central challenge in RL is how toexplore an unknown environment and collect useful feedback efficiently. In recent RL success stories (e.g., super-human performance on video games [Mnih et al., 2015]), we noticethat most of them rely on random exploration strategies, such as -greedy. Similarly, policy gradient method such as REINFORCE [Williams, 1992], perform exploration by injecting randomness into action space and hope the randomness can lead to a good sequence of actions that achieves high total reward. The theoretical RL literature has developed more sophisticated algorithms for efficient exploration (e.g., [Azar et al., 2017]), however, the samplecomplexity of these near-optimal algorithms has to scale exponentially with respect to key parameters of underlying systems such as dimensions of state and action space. Such exponential dependence prohibits a direct application of these theoretically elegant RL algorithms to large-scale applications. In summary, without any further assumptions, RL is hard, both in practice and in theory. In this thesis, we attempt to gain purchase on the RL problem by introducing additional assumptions and sources of information.The first contribution of this thesis comes from improving RL sample complexity via imitation learning. Via leveraging expert’s demonstrations, imitation learning significantly simplifies the tasks of exploration. We consider two settings in this thesis: interactive imitation learning setting where an expert is available to query during training time, and the setting of imitation learning from observation alone, where we only have a set of demonstrations that consist of observations of the expert’s states (no expert actions are recorded). We study in both theory and in practice how one can imitateexperts to reduce sample complexity compared to a pure RL approach. The second contribution comes from model-free Reinforcement Learning. Specifically, we study policyevaluation by building a general reduction from policy evaluation to no-regret online learning which is an active research area that has well-established theoretical foundation. Such a reduction creates a new family of algorithms for provably correct policy evaluation undervery weak assumptions on the generating process. We then provide a through theoretical study and empirical study of two model-free exploration strategies: exploration in actionspace and exploration in parameter space. The third contribution of this work comes from model-based Reinforcement Learning. We provide the first exponential sample complexity separation between model-based RL and general model-free RL approaches. We then provide PAC model-based RL algorithm that can achieve sample efficiency simultaneously for many interesting MDPs such as tabular MDPs, Factored MDPs, Lipschitz continuousMDPs, low rank MDPs, and Linear Quadratic Control. We also provide a more practical model-based RL framework, called Dual Policy Iteration (DPI), via integrating optimalcontrol, model learning, and imitation learning together. Furthermore, we show a general convergence analysis that extends the existing approximate policy iteration theories to DPI. DPI generalizes and provides the first theoretical foundation for recent successful practical RL algorithms such as ExIt and AlphaGo Zero [Anthony et al., 2017, Silver et al., 2017], and provides a theoretical sound and practically efficient way of unifying model-based andmodel-free RL approaches.