Design and training of deep reinforcement learning agents

Abstract

Deep reinforcement learning is a field of research at the intersection of reinforcement learning and deep learning. On one side, the problem that researchers address is the one of reinforcement learning: to act efficiently. A large number of algorithms were developed decades ago in this field to update value functions and policies, explore, and plan. On the other side, deep learning methods provide powerful function approximators to address the problem of representing functions such as policies, value functions, and models. The combination of ideas from these two fields offers exciting new perspectives. However, building successful deep reinforcement learning experiments is particularly difficult due to the large number of elements that must be combined and adjusted appropriately. This thesis proposes a broad overview of the organization of these elements around three main axes: agent design, environment design, and infrastructure design. Arguably, the success of deep reinforcement learning research is due to the tremendous amount of effort that went into each of them, both from a scientific and engineering perspective, and their diffusion via open source repositories. For each of these three axes, a dedicated part of the thesis describes a number of related works that were carried out during the doctoral research.

The first part, devoted to the design of agents, presents two works. The first one addresses the problem of applying discrete action methods to large multidimensional action spaces. A general method called action branching is proposed, and its effectiveness is demonstrated with a novel agent, named BDQ, applied to discretized continuous action spaces. The second work deals with the problem of maximizing the utility of a single transition when learning to achieve a large number of goals. In particular, it focuses on learning to reach spatial locations in games and proposes a new method called Q-map to do so efficiently. An exploration mechanism based on this method is then used to demonstrate the effectiveness of goal-directed exploration. Elements of these works cover some of the main building blocks of agents: update methods, neural architectures, exploration strategies, replays, and hierarchy.

The second part, devoted to the design of environments, also presents two works. The first one shows how various tasks and demonstrations can be combined to learn complex skill spaces that can then be reused to solve even more challenging tasks. The proposed method, called CoMic, extends previous work on motor primitives by using a single multi-clip motion capture tracking task in conjunction with complementary tasks targeting out-of-distribution movements. The second work addresses a particular type of control method vastly neglected in traditional environments but essential for animals: muscle control. An open source codebase called OstrichRL is proposed, containing a musculoskeletal model of an ostrich, an ensemble of tasks, and motion capture data. The results obtained by training a state-of-the-art agent on the proposed tasks show that controlling such a complex system is very difficult and illustrate the importance of using motion capture data. Elements of these works demonstrate the meticulous work that must go into designing environment parts such as: models, observations, rewards, terminations, resets, steps, and demonstrations.

The third part, on the design of infrastructures, presents three works. The first one explains the difference between the types of time limits commonly used in reinforcement learning and why they are often treated inappropriately. In one case, tasks are time-limited by nature and a notion of time should be available to agents to maintain the Markov property of the underlying decision process. In the other case, tasks are not time-limited by nature, but time limits are used for convenience to diversify experiences. This is the most common case. It requires a distinction between time limits and environmental terminations, and bootstrapping should be performed at the end of partial episodes. The second work proposes to unify the most popular deep learning frameworks using a single library called Ivy, and provides new differentiable and framework-agnostic libraries built with it. Four such code bases are provided for gradient-based robot motion planning, mechanics, 3D vision, and differentiable continuous control environments. Finally, the third paper proposes a novel deep reinforcement learning library, called Tonic, built with simplicity and modularity in mind, to accelerate prototyping and evaluation. In particular, it contains implementations of several continuous control agents and a large-scale benchmark. Elements of these works illustrate the different components to consider when building the infrastructure for an experiment: deep learning framework, schedules, and distributed training. Added to these are the various ways to perform evaluations and analyze results for meaningful, interpretable, and reproducible deep reinforcement learning research.