A machine learning (ML) system must learn not only to match the output of a target function on a training set, but also to generalize to novel situations in order to yield accurate predictions at deployment. In most practical applications, the user cannot exhaustively enumerate every possible input to the model; strong generalization performance is therefore crucial to the development of ML systems which are performant and reliable enough to be deployed in the real world. While generalization is well-understood theoretically in a number of hypothesis classes, the impressive generalization performance of deep neural networks has stymied theoreticians. In deep reinforcement learning (RL), our understanding of generalization is further complicated by the conflict between generalization and stability in widely-used RL algorithms. This thesis will provide insight into generalization by studying the learning dynamics of deep neural networks in both supervised and reinforcement learning tasks.

Recently, improving the robustness of policies across different environments attracts increasing attention in the reinforcement learning (RL) community. Existing robust RL methods mostly aim to achieve the max-min robustness by optimizing the policy's performance in the worst-case environment. However, in practice, a user that uses an RL policy may have different preferences over its performance across environments. Clearly, the aforementioned max-min robustness is oftentimes too conservative to satisfy user preference. Therefore, in this paper, we integrate user preference into policy learning in robust RL, and propose a novel User-Oriented Robust RL (UOR-RL) framework. Specifically, we define a new User-Oriented Robustness (UOR) metric for RL, which allocates different weights to the environments according to user preference and generalizes the max-min robustness metric. To optimize the UOR metric, we develop two different UOR-RL training algorithms for the scenarios with or without a priori known environment distribution, respectively. Theoretically, we prove that our UOR-RL training algorithms converge to near-optimal policies even with inaccurate or completely no knowledge about the environment distribution. Furthermore, we carry out extensive experimental evaluations in 4 MuJoCo tasks. The experimental results demonstrate that UOR-RL is comparable to the state-of-the-art baselines under the average and worst-case performance metrics, and more importantly establishes new state-of-the-art performance under the UOR metric.

We revisit the domain of off-policy policy optimization in RL from the perspective of coordinate ascent. One commonly-used approach is to leverage the off-policy policy gradient to optimize a surrogate objective -- the total discounted in expectation return of the target policy with respect to the state distribution of the behavior policy. However, this approach has been shown to suffer from the distribution mismatch issue, and therefore significant efforts are needed for correcting this mismatch either via state distribution correction or a counterfactual method. In this paper, we rethink off-policy learning via Coordinate Ascent Policy Optimization (CAPO), an off-policy actor-critic algorithm that decouples policy improvement from the state distribution of the behavior policy without using the policy gradient. This design obviates the need for distribution correction or importance sampling in the policy improvement step of off-policy policy gradient. We establish the global convergence of CAPO with general coordinate selection and then further quantify the convergence rates of several instances of CAPO with popular coordinate selection rules, including the cyclic and the randomized variants of CAPO. We then extend CAPO to neural policies for a more practical implementation. Through experiments, we demonstrate that CAPO provides a competitive approach to RL in practice.

In classic reinforcement learning algorithms, agents make decisions at discrete and fixed time intervals. The physical duration between one decision and the next becomes a critical hyperparameter. When this duration is too short, the agent needs to make many decisions to achieve its goal, aggravating the problem's difficulty. But when this duration is too long, the agent becomes incapable of controlling the system. Physical systems, however, do not need a constant control frequency. For learning agents, it is desirable to operate with low frequency when possible and high frequency when necessary. We propose a framework called Continuous-Time Continuous-Options (CTCO), where the agent chooses options as sub-policies of variable durations. Such options are time-continuous and can interact with the system at any desired frequency providing a smooth change of actions. The empirical analysis shows that our algorithm is competitive w.r.t. other time-abstraction techniques, such as classic option learning and action repetition, and practically overcomes the difficult choice of the decision frequency.

A central problem in computational biophysics is protein structure prediction, i.e., finding the optimal folding of a given amino acid sequence. This problem has been studied in a classical abstract model, the HP model, where the protein is modeled as a sequence of H (hydrophobic) and P (polar) amino acids on a lattice. The objective is to find conformations maximizing H-H contacts. It is known that even in this reduced setting, the problem is intractable (NP-hard). In this work, we apply deep reinforcement learning (DRL) to the two-dimensional HP model. We can obtain the conformations of best known energies for benchmark HP sequences with lengths from 20 to 50. Our DRL is based on a deep Q-network (DQN). We find that a DQN based on long short-term memory (LSTM) architecture greatly enhances the RL learning ability and significantly improves the search process. DRL can sample the state space efficiently, without the need of manual heuristics. Experimentally we show that it can find multiple distinct best-known solutions per trial. This study demonstrates the effectiveness of deep reinforcement learning in the HP model for protein folding.

As the demand for autonomous driving increases, it is paramount to ensure safety. Early accident prediction using deep learning methods for driving safety has recently gained much attention. In this task, early accident prediction and a point prediction of where the drivers should look are determined, with the dashcam video as input. We propose to exploit the double actors and regularized critics (DARC) method, for the first time, on this accident forecasting platform. We derive inspiration from DARC since it is currently a state-of-the-art reinforcement learning (RL) model on continuous action space suitable for accident anticipation. Results show that by utilizing DARC, we can make predictions 5\% earlier on average while improving in multiple metrics of precision compared to existing methods. The results imply that using our RL-based problem formulation could significantly increase the safety of autonomous driving.

Reinforcement Learning (RL) agents are typically required to operate in dynamic environments, and must develop an ability to quickly adapt to unexpected perturbations in their environment. Promoting this ability is hard, even in single-agent settings Padakandla (2020). For a group this is even more challenging; in addition to the dynamic nature of the environment, agents need to deal with high variance caused by changes in the behavior of other agents. Unsurprisingly, many recent Multi-Agent RL (MARL) works have shown the beneficial effect collaboration between agents has on their performance Xu, Rao, and Bu (2012); Foerster et al. (2016); Lowe et al. (2017); Qian et al. (2019); Jaques et al. (2019); Christianos, Schäfer, and Albrecht (2020). Our objective is to highlight the relationship between a group's ability to collaborate effectively and the group's resilience, which we measure as the group's ability to adapt to perturbations in the environment. Thus, agents that collaborate not only increase their expected utility in a given environment, but are also able to recover a larger fraction of the previous performance after a perturbation occurs. Contrary to investigations of transfer learning Zhu, Lin, and Zhou (2020); Liang and Li (2020) or curriculum learning Portelas et al. (2020), we do not have a stationary target domain in which

We consider a sequential decision making task where we are not allowed to evaluate parameters that violate an a priori unknown (safety) constraint. A common approach is to place a Gaussian process prior on the unknown constraint and allow evaluations only in regions that are safe with high probability. Most current methods rely on a discretization of the domain and cannot be directly extended to the continuous case. Moreover, the way in which they exploit regularity assumptions about the constraint introduces an additional critical hyperparameter. In this paper, we propose an information-theoretic safe exploration criterion that directly exploits the GP posterior to identify the most informative safe parameters to evaluate. Our approach is naturally applicable to continuous domains and does not require additional hyperparameters. We theoretically analyze the method and show that we do not violate the safety constraint with high probability and that we explore by learning about the constraint up to arbitrary precision. Empirical evaluations demonstrate improved data-efficiency and scalability.

Note: This is paper is superseded by the full version (Carroll et al., 2022). Randomly masking and predicting word tokens has been a successful approach in pre-training language models for a variety of downstream tasks. In this work, we observe that the same idea also applies naturally to sequential decision making, where many well-studied tasks like behavior cloning, offline RL, inverse dynamics, and waypoint conditioning correspond to different sequence maskings over a sequence of states, actions, and returns. We introduce the FlexiBiT framework, which provides a unified way to specify models which can be trained on many different sequential decision making tasks. We show that a single FlexiBiT model is simultaneously capable of carrying out many tasks with performance similar to or better than specialized models. Additionally, we show that performance can be further improved by fine-tuning our general model on specific tasks of interest. Masked language modeling (Devlin et al., 2018) is a key technique in natural language processing (NLP). Under this paradigm, models are trained to predict randomly-masked subsets of tokens in a sequence.

We present a new method for generating controllable, dynamically responsive, and photorealistic human animations. Given an image of a person, our system allows the user to generate Physically plausible Upper Body Animation (PUBA) using interaction in the image space, such as dragging their hand to various locations. We formulate a reinforcement learning problem to train a dynamic model that predicts the person's next 2D state (i.e., keypoints on the image) conditioned on a 3D action (i.e., joint torque), and a policy that outputs optimal actions to control the person to achieve desired goals. The dynamic model leverages the expressiveness of 3D simulation and the visual realism of 2D videos. PUBA generates 2D keypoint sequences that achieve task goals while being responsive to forceful perturbation. The sequences of keypoints are then translated by a pose-to-image generator to produce the final photorealistic video.