In this paper, we revisit and improve the convergence of policy gradient (PG), natural PG (NPG) methods, and their variance-reduced variants, under general smooth policy parametrizations. More specifically, with the Fisher information matrix of the policy being positive definite: i) we show that a state-of-the-art variance-reduced PG method, which has only been shown to converge to stationary points, converges to the globally optimal value up to some inherent function approximation error due to policy parametrization; ii) we show that NPG enjoys a lower sample complexity; iii) we propose SRVR-NPG, which incorporates variance-reduction into the NPG update. Our improvements follow from an observation that the convergence of (variance-reduced) PG and NPG methods can improve each other: the stationary convergence analysis of PG can be applied to NPG as well, and the global convergence analysis of NPG can help to establish the global convergence of (variance-reduced) PG methods. Our analysis carefully integrates the advantages of these two lines of works. Thanks to this improvement, we have also made variance-reduction for NPG possible, with both global convergence and an efficient finite-sample complexity.

Developing interactive software, such as websites or games, is a particularly engaging way to learn computer science. However, teaching and giving feedback on such software is time-consuming -- standard approaches require instructors to manually grade student-implemented interactive programs. As a result, online platforms that serve millions, like Code.org, are unable to provide any feedback on assignments for implementing interactive programs, which critically hinders students' ability to learn. One approach toward automatic grading is to learn an agent that interacts with a student's program and explores states indicative of errors via reinforcement learning. However, existing work on this approach only provides binary feedback of whether a program is correct or not, while students require finer-grained feedback on the specific errors in their programs to understand their mistakes. In this work, we show that exploring to discover errors can be cast as a meta-exploration problem. This enables us to construct a principled objective for discovering errors and an algorithm for optimizing this objective, which provides fine-grained feedback. We evaluate our approach on a set of over 700K real anonymized student programs from a Code.org interactive assignment. Our approach provides feedback with 94.3% accuracy, improving over existing approaches by 17.7% and coming within 1.5% of human-level accuracy. Project web page: https://ezliu.github.io/dreamgrader.

The active inference framework (AIF) is a promising new computational framework grounded in contemporary neuroscience that can produce human-like behavior through reward-based learning. In this study, we test the ability for the AIF to capture the role of anticipation in the visual guidance of action in humans through the systematic investigation of a visual-motor task that has been well-explored -- that of intercepting a target moving over a ground plane. Previous research demonstrated that humans performing this task resorted to anticipatory changes in speed intended to compensate for semi-predictable changes in target speed later in the approach. To capture this behavior, our proposed "neural" AIF agent uses artificial neural networks to select actions on the basis of a very short term prediction of the information about the task environment that these actions would reveal along with a long-term estimate of the resulting cumulative expected free energy. Systematic variation revealed that anticipatory behavior emerged only when required by limitations on the agent's movement capabilities, and only when the agent was able to estimate accumulated free energy over sufficiently long durations into the future. In addition, we present a novel formulation of the prior function that maps a multi-dimensional world-state to a uni-dimensional distribution of free-energy. Together, these results demonstrate the use of AIF as a plausible model of anticipatory visually guided behavior in humans.

Traffic signal control affects the daily life of people living in urban areas. The existing system relies on a theory- or rule-based controller in charge of altering the traffic lights based on traffic conditions. The objective is to reduce vehicle delay during unsaturated traffic conditions and maximize the vehicle throughput during congestion. However, the existing traffic signal controller cannot fulfill such objectives, and a human controller can only manage a few intersections. In view of this, recent advancements in artificial intelligence have focused on enabling alternate ways of traffic signal control. Current research on this front has explored reinforcement learning (RL) algorithms as a possible approach.

In this article, we aim to provide a literature review of different formulations and approaches to continual reinforcement learning (RL), also known as lifelong or non-stationary RL. We begin by discussing our perspective on why RL is a natural fit for studying continual learning. We then provide a taxonomy of different continual RL formulations by mathematically characterizing two key properties of non-stationarity, namely, the scope and driver non-stationarity. This offers a unified view of various formulations. Next, we review and present a taxonomy of continual RL approaches. We go on to discuss evaluation of continual RL agents, providing an overview of benchmarks used in the literature and important metrics for understanding agent performance. Finally, we highlight open problems and challenges in bridging the gap between the current state of continual RL and findings in neuroscience. While still in its early days, the study of continual RL has the promise to develop better incremental reinforcement learners that can function in increasingly realistic applications where non-stationarity plays a vital role. These include applications such as those in the fields of healthcare, education, logistics, and robotics.

Offline reinforcement learning (RL) defines a sample-efficient learning paradigm, where a policy is learned from static and previously collected datasets without additional interaction with the environment. The major obstacle to offline RL is the estimation error arising from evaluating the value of out-of-distribution actions. To tackle this problem, most existing offline RL methods attempt to acquire a policy both ``close" to the behaviors contained in the dataset and sufficiently improved over them, which requires a trade-off between two possibly conflicting targets. In this paper, we propose a novel approach, which we refer to as adaptive behavior regularization (ABR), to balance this critical trade-off. By simply utilizing a sample-based regularization, ABR enables the policy to adaptively adjust its optimization objective between cloning and improving over the policy used to generate the dataset. In the evaluation on D4RL datasets, a widely adopted benchmark for offline reinforcement learning, ABR can achieve improved or competitive performance compared to existing state-of-the-art algorithms.

The recent proliferation of the research on multi-agent deep reinforcement learning (MDRL) offers an encouraging way to coordinate multiple connected and automated vehicles (CAVs) to pass the intersection. In this paper, we apply a value decomposition-based MDRL approach (QMIX) to control various CAVs in mixed-autonomy traffic of different densities to efficiently and safely pass the non-signalized intersection with fairish fuel consumption. Implementation tricks including network-level improvements, Q value update by TD ($\lambda$), and reward clipping operation are added to the pure QMIX framework, which is expected to improve the convergence speed and the asymptotic performance of the original version. The efficacy of our approach is demonstrated by several evaluation metrics: average speed, the number of collisions, and average fuel consumption per episode. The experimental results show that our approach's convergence speed and asymptotic performance can exceed that of the original QMIX and the proximal policy optimization (PPO), a state-of-the-art reinforcement learning baseline applied to the non-signalized intersection. Moreover, CAVs under the lower traffic flow controlled by our method can improve their average speed without collisions and consume the least fuel. The training is additionally conducted under the doubled traffic density, where the learning reward converges. Consequently, the model with maximal reward and minimum crashes can still guarantee low fuel consumption, but slightly reduce the efficiency of vehicles and induce more collisions than the lower-traffic counterpart, implying the difficulty of generalizing RL policy to more advanced scenarios.

Humans tame the complexity of mathematical reasoning by developing hierarchies of abstractions. With proper abstractions, solutions to hard problems can be expressed concisely, thus making them more likely to be found. In this paper, we propose Learning Mathematical Abstractions (LEMMA): an algorithm that implements this idea for reinforcement learning agents in mathematical domains. LEMMA augments Expert Iteration with an abstraction step, where solutions found so far are revisited and rewritten in terms of new higher-level actions, which then become available to solve new problems. We evaluate LEMMA on two mathematical reasoning tasks--equation solving and fraction simplification--in a step-by-step fashion. In these two domains, LEMMA improves the ability of an existing agent, both solving more problems and generalizing more effectively to harder problems than those seen during training.

Training novice users to operate an excavator for learning different skills requires the presence of expert teachers. Considering the complexity of the problem, it is comparatively expensive to find skilled experts as the process is time-consuming and requires precise focus. Moreover, since humans tend to be biased, the evaluation process is noisy and will lead to high variance in the final score of different operators with similar skills. In this work, we address these issues and propose a novel strategy for the automatic evaluation of excavator operators. We take into account the internal dynamics of the excavator and the safety criterion at every time step to evaluate the performance. To further validate our approach, we use this score prediction model as a source of reward for a reinforcement learning agent to learn the task of maneuvering an excavator in a simulated environment that closely replicates the real-world dynamics. For a policy learned using these external reward prediction models, our results demonstrate safer solutions following the required dynamic constraints when compared to policy trained with task-based reward functions only, making it one step closer to real-life adoption. For future research, we release our codebase at https://github.com/pranavAL/InvRL_Auto-Evaluate and video results https://drive.google.com/file/d/1jR1otOAu8zrY8mkhUOUZW9jkBOAKK71Z/view?usp=share_link .

While significant research advances have been made in the field of deep reinforcement learning, there have been no concrete adversarial attack strategies in literature tailored for studying the vulnerability of deep reinforcement learning algorithms to membership inference attacks. In such attacking systems, the adversary targets the set of collected input data on which the deep reinforcement learning algorithm has been trained. To address this gap, we propose an adversarial attack framework designed for testing the vulnerability of a state-of-the-art deep reinforcement learning algorithm to a membership inference attack. In particular, we design a series of experiments to investigate the impact of temporal correlation, which naturally exists in reinforcement learning training data, on the probability of information leakage. Moreover, we compare the performance of \emph{collective} and \emph{individual} membership attacks against the deep reinforcement learning algorithm. Experimental results show that the proposed adversarial attack framework is surprisingly effective at inferring data with an accuracy exceeding $84\%$ in individual and $97\%$ in collective modes in three different continuous control Mujoco tasks, which raises serious privacy concerns in this regard. Finally, we show that the learning state of the reinforcement learning algorithm influences the level of privacy breaches significantly.