In this paper we investigate the properties of representations learned by deep reinforcement learning systems. Much of the early work on representations for reinforcement learning focused on designing fixed-basis architectures to achieve properties thought to be desirable, such as orthogonality and sparsity. In contrast, the idea behind deep reinforcement learning methods is that the agent designer should not encode representational properties, but rather that the data stream should determine the properties of the representation -- good representations emerge under appropriate training schemes. In this paper we bring these two perspectives together, empirically investigating the properties of representations that support transfer in reinforcement learning. We introduce and measure six representational properties over more than 25 thousand agent-task settings. We consider Deep Q-learning agents with different auxiliary losses in a pixel-based navigation environment, with source and transfer tasks corresponding to different goal locations. We develop a method to better understand why some representations work better for transfer, through a systematic approach varying task similarity and measuring and correlating representation properties with transfer performance. We demonstrate the generality of the methodology by investigating representations learned by a Rainbow agent that successfully transfer across games modes in Atari 2600.

In our recent AAAI 2023 paper, Misspecification in Inverse Reinforcement Learning (Skalse and Abate, 2023), we study the question of how robust the inverse reinforcement learning problem is to misspecification of the underlying behavioural model (namely, how the agent's preferences relate to its behaviour). We provide a mathematical framework for reasoning about this question, and use that framework (based on equivalence classes and orders) to derive necessary and sufficient conditions describing what types of misspecification each of the standard behavioural models are (or are not) robust to. Moreover, we provide several results and formal tools, which can be used to study the misspecification robustness of any behavioural models that may be newly developed. Below, we will first explain the motivation for this work. Then, we will explain our results and, finally, describe ways to extend them. Inverse reinforcement learning (IRL) is an area of machine learning concerned with inferring what objective an agent is pursuing, based on the actions taken by that agent.

While off-policy reinforcement learning (RL) algorithms are sample efficient due to gradient-based updates and data reuse in the replay buffer, they struggle with convergence to local optima due to limited exploration. On the other hand, population-based algorithms offer a natural exploration strategy, but their heuristic black-box operators are inefficient. Recent algorithms have integrated these two methods, connecting them through a shared replay buffer. However, the effect of using diverse data from population optimization iterations on off-policy RL algorithms has not been thoroughly investigated. In this paper, we first analyze the use of off-policy RL algorithms in combination with population-based algorithms, showing that the use of population data could introduce an overlooked error and harm performance. To test this, we propose a uniform and scalable training design and conduct experiments on our tailored framework in robot locomotion tasks from the OpenAI gym. Our results substantiate that using population data in off-policy RL can cause instability during training and even degrade performance. To remedy this issue, we further propose a double replay buffer design that provides more on-policy data and show its effectiveness through experiments. Our results offer practical insights for training these hybrid methods.

The skill of pivoting an object with a robotic system is challenging for the external forces that act on the system, mainly given by contact interaction. The complexity increases when the same skills are required to generalize across different objects. This paper proposes a framework for learning robust and generalizable pivoting skills, which consists of three steps. First, we learn a pivoting policy on an ``unitary'' object using Reinforcement Learning (RL). Then, we obtain the object's feature space by supervised learning to encode the kinematic properties of arbitrary objects. Finally, to adapt the unitary policy to multiple objects, we learn data-driven projections based on the object features to adjust the state and action space of the new pivoting task. The proposed approach is entirely trained in simulation. It requires only one depth image of the object and can zero-shot transfer to real-world objects. We demonstrate robustness to sim-to-real transfer and generalization to multiple objects.

Autocurricular training is an important sub-area of multi-agent reinforcement learning~(MARL) that allows multiple agents to learn emergent skills in an unsupervised co-evolving scheme. The robotics community has experimented autocurricular training with physically grounded problems, such as robust control and interactive manipulation tasks. However, the asymmetric nature of these tasks makes the generation of sophisticated policies challenging. Indeed, the asymmetry in the environment may implicitly or explicitly provide an advantage to a subset of agents which could, in turn, lead to a low-quality equilibrium. This paper proposes a novel game-theoretic algorithm, Stackelberg Multi-Agent Deep Deterministic Policy Gradient (ST-MADDPG), which formulates a two-player MARL problem as a Stackelberg game with one player as the `leader' and the other as the `follower' in a hierarchical interaction structure wherein the leader has an advantage. We first demonstrate that the leader's advantage from ST-MADDPG can be used to alleviate the inherent asymmetry in the environment. By exploiting the leader's advantage, ST-MADDPG improves the quality of a co-evolution process and results in more sophisticated and complex strategies that work well even against an unseen strong opponent.

Robots are more capable of achieving manipulation tasks for everyday activities than before. But the safety of manipulation skills that robots employ is still an open problem. Considering all possible failures during skill learning increases the complexity of the process and restrains learning an optimal policy. Beyond that, in unstructured environments, it is not easy to enumerate all possible failures beforehand. In the context of safe skill manipulation, we reformulate skills as base and failure prevention skills where base skills aim at completing tasks and failure prevention skills focus on reducing the risk of failures to occur. Then, we propose a modular and hierarchical method for safe robot manipulation by augmenting base skills by learning failure prevention skills with reinforcement learning, forming a skill library to address different safety risks. Furthermore, a skill selection policy that considers estimated risks is used for the robot to select the best control policy for safe manipulation. Our experiments show that the proposed method achieves the given goal while ensuring safety by preventing failures. We also show that with the proposed method, skill learning is feasible, novel failures are easily adaptable, and our safe manipulation tools can be transferred to the real environment.

Federated learning is an approach wherein clients learn collaboratively by sharing their locally trained models (not their data) with a federating agent, which periodically combines their models and returns the federated model to the clients for further refinement (Kairouz et al., 2021; Wang et al., 2021). Federated learning has seen much success in supervised learning applications due to its ability to generate well-trained models using small amounts of data at each client while preserving privacy and reducing the usage of communication resources. Recently, there is a growing interest in employing federated learning for online RL problems where each client collects data online by following its own Markovian trajectory, while simultaneously updating the model parameters (Khodadadian et al., 2022; Nadiger et al., 2019; Qi et al., 2021). However, such an online learning approach requires sequential interactions with the environment or the simulator, which may not be feasible in many real-world applications. Instead, each clients may have pre-collected operational data generated according to a client-specific behavior policy. The federated offline reinforcement learning problem is to learn the optimal policy using these heterogeneous offline data sets distributed across the clients and collected by different unknown behavior policies, without sharing the data explicitly. The framework of offline reinforcement learning (Levine et al., 2020) offers a way to learn the policy only using the offline data collected according a behavior policy, without any direct interactions with the environment. However, naively combining an off-the-shelf offline RL algorithm such as TD3-BC (Fujimoto & Gu, 2021) with an off-the-shelf federated supervised learning approach such as FedAvg (McMahan et al., 2017) will lead to a poorly performing policy, as we show later (see Figure 1-3). Federated offline RL is significantly more challenging than its supervised learning counterpart and the centralized offline RL because of the following reasons.

We introduce Masked Trajectory Models (MTM) as a generic abstraction for sequential decision making. MTM takes a trajectory, such as a state-action sequence, and aims to reconstruct the trajectory conditioned on random subsets of the same trajectory. By training with a highly randomized masking pattern, MTM learns versatile networks that can take on different roles or capabilities, by simply choosing appropriate masks at inference time. For example, the same MTM network can be used as a forward dynamics model, inverse dynamics model, or even an offline RL agent. Through extensive experiments in several continuous control tasks, we show that the same MTM network -- i.e. same weights -- can match or outperform specialized networks trained for the aforementioned capabilities. Additionally, we find that state representations learned by MTM can significantly accelerate the learning speed of traditional RL algorithms. Finally, in offline RL benchmarks, we find that MTM is competitive with specialized offline RL algorithms, despite MTM being a generic self-supervised learning method without any explicit RL components. Code is available at https://github.com/facebookresearch/mtm

Artificial neural networks (ANNs) are increasingly used as research models, but questions remain about their generalizability and representational invariance. Biological neural networks under social constraints evolved to enable communicable representations, demonstrating generalization capabilities. This study proposes a communication protocol between cooperative agents to analyze the formation of individual and shared abstractions and their impact on task performance. This communication protocol aims to mimic language features by encoding high-dimensional information through low-dimensional representation. Using grid-world mazes and reinforcement learning, teacher ANNs pass a compressed message to a student ANN for better task completion. Through this, the student achieves a higher goal-finding rate and generalizes the goal location across task worlds. Further optimizing message content to maximize student reward improves information encoding, suggesting that an accurate representation in the space of messages requires bi-directional input. This highlights the role of language as a common representation between agents and its implications on generalization capabilities.

Training a dialogue policy using deep reinforcement learning requires a lot of exploration of the environment. The amount of wasted invalid exploration makes their learning inefficient. In this paper, we find and define an important reason for the invalid exploration: dead-ends. When a conversation enters a dead-end state, regardless of the actions taken afterward, it will continue in a dead-end trajectory until the agent reaches a termination state or maximum turn. We propose a dead-end resurrection (DDR) algorithm that detects the initial dead-end state in a timely and efficient manner and provides a rescue action to guide and correct the exploration direction. To prevent dialogue policies from repeatedly making the same mistake, DDR also performs dialogue data augmentation by adding relevant experiences containing dead-end states. We first validate the dead-end detection reliability and then demonstrate the effectiveness and generality of the method by reporting experimental results on several dialogue datasets from different domains.