Efficient Deep Reinforcement Learning via Planning, Generalization, and Improved Exploration

Abstract

Reinforcement learning (RL) is a general-purpose machine learning framework, which considers an agent that makes sequential decisions in an environment to maximize its reward. Deep reinforcement learning (DRL) approaches use deep neural networks as non-linear function approximators that parameterize policies or value functions directly from raw observations in RL. Although DRL approaches have been shown to be successful on many challenging RL benchmarks, much of the prior work has mainly focused on learning a single task in a model-free setting, which is often sample-inefficient. On the other hand, humans have abilities to acquire knowledge by learning a model of the world in an unsupervised fashion, use such knowledge to plan ahead for decision making, transfer knowledge between many tasks, and generalize to previously unseen circumstances from the pre-learned knowledge. Developing such abilities are some of the fundamental challenges for building RL agents that can learn as efficiently as humans.

As a step towards developing the aforementioned capabilities in RL, this thesis develops new DRL techniques to address three important challenges in RL: 1) planning via prediction, 2) rapidly generalizing to new environments and tasks, and 3) efficient exploration in complex environments.

The first part of the thesis discusses how to learn a dynamics model of the environment using deep neural networks and how to use such a model for planning in complex domains where observations are high-dimensional. Specifically, we present neural network architectures for action-conditional video prediction and demonstrate improved exploration in RL. In addition, we present a neural network architecture that performs lookahead planning by predicting the future only in terms of rewards and values without predicting observations. We then discuss why this approach is beneficial compared to conventional model-based planning approaches.

The second part of the thesis considers generalization to unseen environments and tasks. We first introduce a set of cognitive tasks in a 3D environment and present memory-based DRL architectures that generalize better to previously unseen 3D environments compared to existing baselines. In addition, we introduce a new multi-task RL problem where the agent should learn to execute different tasks depending on given instructions and generalize to new instructions in a zero-shot fashion. We present a new hierarchical DRL architecture that learns to generalize over previously unseen task descriptions with minimal prior knowledge.

The third part of the thesis discusses how exploiting past experiences can indirectly drive deep exploration and improve sample-efficiency. In particular, we propose a new off-policy learning algorithm, called self-imitation learning, which learns a policy to reproduce past good experiences. We empirically show that self-imitation learning indirectly encourages the agent to explore reasonably good state spaces and thus significantly improves sample-efficiency on RL domains where exploration is challenging.

Overall, the main contribution of this thesis are to explore several fundamental challenges in RL in the context of DRL and develop new DRL architectures and algorithms to address such challenges. This allows us to understand how deep learning can be used to improve sample efficiency, and thus come closer to human-like learning abilities.