Search

Recent developments in Deep Reinforcement Learning (DRL) have shown tremendous progress in robotics control, Atari games, board games such as Go, etc. However, model free DRL still has limited use cases due to its poor sampling efficiency and generalization on a variety of tasks. In this thesis, two particular drawbacks of DRL are investigated: 1) the poor generalization abilities of model free DRL. More specifically, how to generalize an agent's policy to unseen environments and generalize to task performance on different data representations (e.g. image based or graph based) 2) The reality gap issue in DRL. That is, how to effectively transfer a policy learned in a simulator to the real world. This thesis makes several novel contributions to the field of DRL which are outlined sequentially in the following. Among these contributions is the generalized value iteration network (GVIN) algorithm, which is an end-to-end neural network planning module extending the work of Value Iteration Networks (VIN). GVIN emulates the value iteration algorithm by using a novel graph convolution operator, which enables GVIN to learn and plan on irregular spatial graphs. Additionally, this thesis proposes three novel, differentiable kernels as graph convolution operators and shows that the embedding-based kernel achieves the best performance. Furthermore, an improvement upon traditional $n$-step $Q$-learning that stabilizes training for VIN and GVIN is demonstrated. Additionally, the equivalence between GVIN and graph neural networks is outlined and shown that GVIN can be further extended to address both control and inference problems. The final subject which falls under the graph domain that is studied in this thesis is graph embeddings. Specifically, this work studies a general graph embedding framework GEM-F that unifies most of the previous graph embedding algorithms. Based on the contributions made during the analysis of GEM-F, a novel algorithm called WarpMap which outperforms DeepWalk and node2vec in the unsupervised learning settings is proposed. The aforementioned reality gap in DRL prohibits a significant portion of research from reaching the real world setting. The latter part of this work studies and analyzes domain transfer techniques in an effort to bridge this gap. Typically, domain transfer in RL consists of representation transfer and policy transfer. In this work, the focus is on representation transfer for vision based applications. More specifically, aligning the feature representation from source domain to target domain in an unsupervised fashion. In this approach, a linear mapping function is considered to fuse modules that are trained in different domains. Proposed are two improved adversarial learning methods to enhance the training quality of the mapping function. Finally, the thesis demonstrates the effectiveness of domain alignment among different weather conditions in the CARLA autonomous driving simulator.