Massively parallel GPU simulation environments have accelerated reinforcement learning (RL) research by enabling fast data collection for on-policy RL algorithms like Proximal Policy Optimization (PPO). To maximize throughput, it is common to use short rollouts per policy update, increasing the update-to-data (UTD) ratio. However, we find that, in this setting, standard synchronous resets introduce harmful nonstationarity, skewing the learning signal and destabilizing training. We introduce staggered resets, a simple yet effective technique where environments are initialized and reset at varied points within the task horizon. This yields training batches with greater temporal diversity, reducing the nonstationarity induced by synchronized rollouts. We characterize dimensions along which RL environments can benefit significantly from staggered resets through illustrative toy environments. We then apply this technique to challenging high-dimensional robotics environments, achieving significantly higher sample efficiency, faster wall-clock convergence, and stronger final performance. Finally, this technique scales better with more parallel environments compared to naive synchronized rollouts.

Intoyenvironments, Attainable Utility Preservation (AUP)avoided side effects by penalizing shifts in the ability to achieve randomly generated goals [22]. We scale this approach to large, randomly generated environments based onConway'sGame ofLife.

Massively parallel GPU simulation environments have accelerated reinforcement learning (RL) research by enabling fast data collection for on-policy RL algorithms like Proximal Policy Optimization (PPO). To maximize throughput, it is common to use short rollouts per policy update, increasing the update-to-data (UTD) ra- tio. However, we find that, in this setting, standard synchronous resets introduce harmful nonstationarity, skewing the learning signal and destabilizing training. We introduce staggered resets, a simple yet effective technique where environments are initialized and reset at varied points within the task horizon. This yields training batches with greater temporal diversity, reducing the nonstationarity induced by synchronized rollouts. We characterize dimensions along which RL environments can benefit significantly from staggered resets through illustrative toy environ- ments. We then apply this technique to challenging high-dimensional robotics environments, achieving significantly higher sample efficiency, faster wall-clock convergence, and stronger final performance. Finally, this technique scales better with more parallel environments compared to naive synchronized rollouts.

We present MDP Playground, a testbed for Reinforcement Learning (RL) agents with dimensions of hardness that can be controlled independently to challenge agents in different ways and obtain varying degrees of hardness in toy and complex RL environments. We consider and allow control over a wide variety of dimensions, including delayed rewards, sequence lengths, reward density, stochasticity, image representations, irrelevant features, time unit, action range and more. We define a parameterised collection of fast-to-run toy environments in OpenAI Gym by varying these dimensions and propose to use these to understand agents better. We then show how to design experiments using MDP Playground to gain insights on the toy environments. We also provide wrappers that can inject many of these dimensions into any Gym environment. We experiment with these wrappers on Atari and Mujoco to allow for understanding the effects of these dimensions on environments that are more complex than the toy environments. We also compare the effect of the dimensions on the toy and complex environments. Finally, we show how to use MDP Playground to debug agents, to study the interaction of multiple dimensions and describe further use-cases.

Recommender systems (RS) mediate human experience online. Most RS act to optimize metrics that are imperfectly aligned with the best-interest of users but are easy to measure, like ad-clicks and user engagement. This has resulted in a host of hard-to-measure side-effects: political polarization, addiction, fake news. RS design faces a recommender alignment problem: that of aligning recommendations with the goals of users, system designers, and society as a whole. But how do we test and compare potential solutions to align RS? Their massive scale makes them costly and risky to test in deployment. We synthesized a simple abstract modelling framework to guide future work. To illustrate it, we construct a toy experiment where we ask: "How can we evaluate the consequences of using user retention as a reward function?" To answer the question, we learn recommender policies that optimize reward functions by controlling graph dynamics on a toy environment. Based on the effects that trained recommenders have on their environment, we conclude that engagement maximizers generally lead to worse outcomes than aligned recommenders but not always. After learning, we examine competition between RS as a potential solution to RS alignment. We find that it generally makes our toy-society better-off than it would be under the absence of recommendation or engagement maximizers. In this work, we aimed for a broad scope, touching superficially on many different points to shed light on how an end-to-end study of reward functions for recommender systems might be done. Recommender alignment is a pressing and important problem. Attempted solutions are sure to have far-reaching impacts. Here, we take a first step in developing methods to evaluating and comparing solutions with respect to their impacts on society.

A major challenge in the Deep RL (DRL) community is to train agents able to generalize their control policy over situations never seen in training. Training on diverse tasks has been identified as a key ingredient for good generalization, which pushed researchers towards using rich procedural task generation systems controlled through complex continuous parameter spaces. In such complex task spaces, it is essential to rely on some form of Automatic Curriculum Learning (ACL) to adapt the task sampling distribution to a given learning agent, instead of randomly sampling tasks, as many could end up being either trivial or unfeasible. Since it is hard to get prior knowledge on such task spaces, many ACL algorithms explore the task space to detect progress niches over time, a costly tabula-rasa process that needs to be performed for each new learning agents, although they might have similarities in their capabilities profiles. To address this limitation, we introduce the concept of Meta-ACL, and formalize it in the context of black-box RL learners, i.e. algorithms seeking to generalize curriculum generation to an (unknown) distribution of learners. In this work, we present AGAIN, a first instantiation of Meta-ACL, and showcase its benefits for curriculum generation over classical ACL in multiple simulated environments including procedurally generated parkour environments with learners of varying morphologies. Videos and code are available at https://sites.google.com/view/meta-acl .

Reinforcement learning (RL) algorithms update an agent's parameters according to one of several possible rules, discovered manually through years of research. Automating the discovery of update rules from data could lead to more efficient algorithms, or algorithms that are better adapted to specific environments. Although there have been prior attempts at addressing this significant scientific challenge, it remains an open question whether it is feasible to discover alternatives to fundamental concepts of RL such as value functions and temporal-difference learning. This paper introduces a new meta-learning approach that discovers an entire update rule which includes both 'what to predict' (e.g. value functions) and 'how to learn from it' (e.g. bootstrapping) by interacting with a set of environments. The output of this method is an RL algorithm that we call Learned Policy Gradient (LPG). Empirical results show that our method discovers its own alternative to the concept of value functions. Furthermore it discovers a bootstrapping mechanism to maintain and use its predictions. Surprisingly, when trained solely on toy environments, LPG generalises effectively to complex Atari games and achieves non-trivial performance. This shows the potential to discover general RL algorithms from data.

Recent works in high-dimensional model-predictive control and model-based reinforcement learning with learned dynamics and reward models have resorted to population-based optimization methods, such as the Cross-Entropy Method (CEM), for planning a sequence of actions. To decide on an action to take, CEM conducts a search for the action sequence with the highest return according to the dynamics model and reward. Action sequences are typically randomly sampled from an unconditional Gaussian distribution and evaluated on the environment. This distribution is iteratively updated towards action sequences with higher returns. However, this planning method can be very inefficient, especially for high-dimensional action spaces. An alternative line of approaches optimize action sequences directly via gradient descent, but are prone to local optima. We propose a method to solve this planning problem by interleaving CEM and gradient descent steps in optimizing the action sequence. Our experiments show faster convergence of the proposed hybrid approach, even for high-dimensional action spaces, avoidance of local minima, and better or equal performance to CEM. Code accompanying the paper is available here 1 .