Q-learning has long been one of the most popular reinforcement learning algorithms, and theoretical analysis of Q-learning has been an active research topic for decades. Although researches on asymptotic convergence analysis of Q-learning have a long tradition, non-asymptotic convergence has only recently come under active study. The main goal of this paper is to investigate new finite-time analysis of asynchronous Q-learning under Markovian observation models via a control system viewpoint. In particular, we introduce a discrete-time time-varying switching system model of Q-learning with diminishing step-sizes for our analysis, which significantly improves recent development of the switching system analysis with constant step-sizes, and leads to \(\mathcal{O}\left( \sqrt{\frac{\log k}{k}} \right)\) convergence rate that is comparable to or better than most of the state of the art results in the literature. In the mean while, a technique using the similarly transformation is newly applied to avoid the difficulty in the analysis posed by diminishing step-sizes. The proposed analysis brings in additional insights, covers different scenarios, and provides new simplified templates for analysis to deepen our understanding on Q-learning via its unique connection to discrete-time switching systems.

A hallmark of human intelligence is the ability to construct self-contained chunks of knowledge and reuse them in novel combinations for solving different problems. Learning such compositional structures has been a challenge for artificial systems, due to the underlying combinatorial search. To date, research into compositional learning has largely proceeded separately from work on lifelong or continual learning. This dissertation integrated these two lines of work to present a general-purpose framework for lifelong learning of functionally compositional structures. The framework separates the learning into two stages: learning how to combine existing components to assimilate a novel problem, and learning how to adapt the existing components to accommodate the new problem. This separation explicitly handles the trade-off between stability and flexibility. This dissertation instantiated the framework into various supervised and reinforcement learning (RL) algorithms. Supervised learning evaluations found that 1) compositional models improve lifelong learning of diverse tasks, 2) the multi-stage process permits lifelong learning of compositional knowledge, and 3) the components learned by the framework represent self-contained and reusable functions. Similar RL evaluations demonstrated that 1) algorithms under the framework accelerate the discovery of high-performing policies, and 2) these algorithms retain or improve performance on previously learned tasks. The dissertation extended one lifelong compositional RL algorithm to the nonstationary setting, where the task distribution varies over time, and found that modularity permits individually tracking changes to different elements in the environment. The final contribution of this dissertation was a new benchmark for compositional RL, which exposed that existing methods struggle to discover the compositional properties of the environment.

In decentralized cooperative multi-agent reinforcement learning, agents can aggregate information from one another to learn policies that maximize a team-average objective function. Despite the willingness to cooperate with others, the individual agents may find direct sharing of information about their local state, reward, and value function undesirable due to privacy issues. In this work, we introduce a decentralized actor-critic algorithm with TD error aggregation that does not violate privacy issues and assumes that communication channels are subject to time delays and packet dropouts. The cost we pay for making such weak assumptions is an increased communication burden for every agent as measured by the dimension of the transmitted data. Interestingly, the communication burden is only quadratic in the graph size, which renders the algorithm applicable in large networks. We provide a convergence analysis under diminishing step size to verify that the agents maximize the team-average objective function.

The ability to generalize to previously unseen tasks with little to no supervision is a key challenge in modern machine learning research. It is also a cornerstone of a future "General AI". Any artificially intelligent agent deployed in a real world application, must adapt on the fly to unknown environments. Researchers often rely on reinforcement and imitation learning to provide online adaptation to new tasks, through trial and error learning. However, this can be challenging for complex tasks which require many timesteps or large numbers of subtasks to complete. These "long horizon" tasks suffer from sample inefficiency and can require extremely long training times before the agent can learn to perform the necessary longterm planning. In this work, we introduce CASE which attempts to address these issues by training an Imitation Learning agent using adaptive "near future" subgoals. These subgoals are recalculated at each step using compositional arithmetic in a learned latent representation space. In addition to improving learning efficiency for standard long-term tasks, this approach also makes it possible to perform one-shot generalization to previously unseen tasks, given only a single reference trajectory for the task in a different environment. Our experiments show that the proposed approach consistently outperforms the previous state-of-the-art compositional Imitation Learning approach by 30%.

Reliable navigation systems have a wide range of applications in robotics and autonomous driving. Current approaches employ an open-loop process that converts sensor inputs directly into actions. However, these open-loop schemes are challenging to handle complex and dynamic real-world scenarios due to their poor generalization. Imitating human navigation, we add a reasoning process to convert actions back to internal latent states, forming a two-stage closed loop of perception, decision-making, and reasoning. Firstly, VAE-Enhanced Demonstration Learning endows the model with the understanding of basic navigation rules. Then, two dual processes in RL-Enhanced Interaction Learning generate reward feedback for each other and collectively enhance obstacle avoidance capability. The reasoning model can substantially promote generalization and robustness, and facilitate the deployment of the algorithm to real-world robots without elaborate transfers. Experiments show our method is more adaptable to novel scenarios compared with state-of-the-art approaches.

Despite recent advances in natural language understanding and generation, and decades of research on the development of conversational bots, building automated agents that can carry on rich open-ended conversations with humans "in the wild" remains a formidable challenge. In this work we develop a real-time, open-ended dialogue system that uses reinforcement learning (RL) to power a bot's conversational skill at scale. Our work pairs the succinct embedding of the conversation state generated using SOTA (supervised) language models with RL techniques that are particularly suited to a dynamic action space that changes as the conversation progresses. Trained using crowd-sourced data, our novel system is able to substantially exceeds the (strong) baseline supervised model with respect to several metrics of interest in a live experiment with real users of the Google Assistant.

Many studies have proposed methods for optimizing the dialogue performance of an entire pipeline task-oriented dialogue system by jointly training modules in the system using reinforcement learning. However, these methods are limited in that they can only be applied to modules implemented using trainable neural-based methods. To solve this problem, we propose a method for optimizing a pipeline system composed of modules implemented with arbitrary methods for dialogue performance. With our method, neural-based components called post-processing networks (PPNs) are installed inside such a system to post-process the output of each module. All PPNs are updated to improve the overall dialogue performance of the system by using reinforcement learning, not necessitating each module to be differentiable. Through dialogue simulation and human evaluation on the MultiWOZ dataset, we show that our method can improve the dialogue performance of pipeline systems consisting of various modules.

Any reinforcement learning system must be able to identify which past events contributed to observed outcomes, a problem known as credit assignment. A common solution to this problem is to use an eligibility trace to assign credit to recency-weighted set of experienced events. However, in many realistic tasks, the set of recently experienced events are only one of the many possible action events that could have preceded the current outcome. This suggests that reinforcement learning can be made more efficient by allowing credit assignment to any viable preceding state, rather than only those most recently experienced. Accordingly, we examine ``Predecessor Features'', the fully bootstrapped version of van Hasselt's ``Expected Trace'', an algorithm that achieves this richer form of credit assignment. By maintaining a representation that approximates the expected sum of past occupancies, this algorithm allows temporal difference (TD) errors to be propagated accurately to a larger number of predecessor states than conventional methods, greatly improving learning speed. The algorithm can also be naturally extended from tabular state representation to feature representations allowing for increased performance on a wide range of environments. We demonstrate several use cases for Predecessor Features and compare its performance with other approaches.

Bundle recommender systems recommend sets of items (e.g., pants, shirt, and shoes) to users, but they often suffer from two issues: significant interaction sparsity and a large output space. In this work, we extend multi-round conversational recommendation (MCR) to alleviate these issues. MCR, which uses a conversational paradigm to elicit user interests by asking user preferences on tags (e.g., categories or attributes) and handling user feedback across multiple rounds, is an emerging recommendation setting to acquire user feedback and narrow down the output space, but has not been explored in the context of bundle recommendation. In this work, we propose a novel recommendation task named Bundle MCR. We first propose a new framework to formulate Bundle MCR as Markov Decision Processes (MDPs) with multiple agents, for user modeling, consultation and feedback handling in bundle contexts. Under this framework, we propose a model architecture, called Bundle Bert (Bunt) to (1) recommend items, (2) post questions and (3) manage conversations based on bundle-aware conversation states. Moreover, to train Bunt effectively, we propose a two-stage training strategy. In an offline pre-training stage, Bunt is trained using multiple cloze tasks to mimic bundle interactions in conversations. Then in an online fine-tuning stage, Bunt agents are enhanced by user interactions. Our experiments on multiple offline datasets as well as the human evaluation show the value of extending MCR frameworks to bundle settings and the effectiveness of our Bunt design.

With the continuous growth of the global economy and markets, resource imbalance has risen to be one of the central issues in real logistic scenarios. In marine transportation, this trade imbalance leads to Empty Container Repositioning (ECR) problems. Once the freight has been delivered from an exporting country to an importing one, the laden will turn into empty containers that need to be repositioned to satisfy new goods requests in exporting countries. In such problems, the performance that any cooperative repositioning policy can achieve strictly depends on the routes that vessels will follow (i.e., fleet deployment). Historically, Operation Research (OR) approaches were proposed to jointly optimize the repositioning policy along with the fleet of vessels. However, the stochasticity of future supply and demand of containers, together with black-box and non-linear constraints that are present within the environment, make these approaches unsuitable for these scenarios. In this paper, we introduce a novel framework, Configurable Semi-POMDPs, to model this type of problems. Furthermore, we provide a two-stage learning algorithm, "Configure & Conquer" (CC), that first configures the environment by finding an approximation of the optimal fleet deployment strategy, and then "conquers" it by learning an ECR policy in this tuned environmental setting. We validate our approach in large and real-world instances of the problem. Our experiments highlight that CC avoids the pitfalls of OR methods and that it is successful at optimizing both the ECR policy and the fleet of vessels, leading to superior performance in world trade environments.