In dynamic environments, Q-learning is an automaton that (i) provides estimates (Q-values) of the continuation values associated with each available action; and (ii) follows the naive policy of almost always choosing the action with highest Q-value. We consider a family of automata that are based on Q-values but whose policy may systematically favor some actions over others, for example through a bias that favors cooperation. In the spirit of Compte and Postlewaite [2018], we look for equilibrium biases within this family of Q-based automata. We examine classic games under various monitoring technologies and find that equilibrium biases may strongly foster collusion.

In task-oriented dialogs (TOD), reinforcement learning (RL) algorithms train a model to directly optimize response for task-related metrics. However, RL needs to perform exploration, which can be time-consuming due to the slow auto-regressive sequence generation process. We investigate an approach to create a more efficient RL-based algorithm to improve TOD performance in an offline setting. First, we use a faster generation procedure that samples from independent next-word distributions after training the language model (LM) with supervised learning. We then introduce a fine-grained reward function to help the model focus on learning key information in a dialog, by measuring the importance and semantic closeness of each generated token. Experiments on the MultiWoZ dataset show our new training algorithm, Keywords Reinforcement Learning with Next-word Sampling (KRLS), achieves state-of-the-art performance on the end-to-end response generation task, with a 15% training time reduction compared to a standard RL algorithm using auto-regressive generation.

As two cornerstones of modern day technologies, speech processing and natural language processing (NLP) are innately sequence learning problems to extract information from these linguistic or speech signals and provide insights into interactive systems to communicate in human understandable languages. The sequential and interactive nature of these problems can make them well-suited into the algorithmic framework of reinforcement learning (RL). In a reinforcement learning setting, an agent interacts with an environment through observations and actions, and based on the reward feedback attributed by the underlying reward function of this environment, the agent learns how to perform the task of interest through trials and errors. While the successful applications of reinforcement learning have been highlighted by a wide range of surveys in many real-world engineering domains such as robotics [1], vision [2], finance [3], healthcare [4], linguistics [5], and energy management [6], there have not been one for the rich community of both the speech and language domains. This is the first survey that emphasizes the synergy among the growing fields of the speech processing, natural language processing and the reinforcement learning. We aim to fill this gap by adopting a complete, timely and classical view of the reinforcement learning problems and their connections to speech and language processing.

Offline constrained reinforcement learning (RL) aims to learn a decision making policy that performs well while satisfying safety constraints given a dataset of trajectories collected from historical experiments. It enjoys the benefits of offline RL [22]: not requiring interaction with the environment enables real-world applications where collecting interaction data is expensive (e.g., robotics [18, 23]) or dangerous (e.g., healthcare [30]). It also enjoys the benefits of constrained RL [1]: being able to specify constraints to the behavior of the agent enables real-world applications with safety concerns (e.g., smart grid [31], robotics [14]). Offline constrained RL with function approximation (e.g., neural networks) is of particular interest because function approximation can encode inductive biases to allow sample-efficient learning in large state spaces. As is the case for offline unconstrained RL [29, 32], offline constrained RL with function approximation requires two classes of assumptions for sample-efficient learning.

Embodied agents operate in a structured world, often solving tasks with spatial, temporal, and permutation symmetries. Most algorithms for planning and model-based reinforcement learning (MBRL) do not take this rich geometric structure into account, leading to sample inefficiency and poor generalization. We introduce the Equivariant Diffuser for Generating Interactions (EDGI), an algorithm for MBRL and planning that is equivariant with respect to the product of the spatial symmetry group SE(3), the discrete-time translation group Z, and the object permutation group Sn. EDGI follows the Diffuser framework (Janner et al., 2022) in treating both learning a world model and planning in it as a conditional generative modeling problem, training a diffusion model on an offline trajectory dataset. We introduce a new SE(3)xZxSn-equivariant diffusion model that supports multiple representations. We integrate this model in a planning loop, where conditioning and classifier guidance let us softly break the symmetry for specific tasks as needed. On object manipulation and navigation tasks, EDGI is substantially more sample efficient and generalizes better across the symmetry group than non-equivariant models.

To solve a task with reinforcement learning (RL), it is necessary to formally specify the goal of that task. Although most RL algorithms require that the goal is formalised as a Markovian reward function, alternatives have been developed (such as Linear Temporal Logic and Multi-Objective Reinforcement Learning). Moreover, it is well known that some of these formalisms are able to express certain tasks that other formalisms cannot express. However, there has not yet been any thorough analysis of how these formalisms relate to each other in terms of expressivity. In this work, we fill this gap in the existing literature by providing a comprehensive comparison of the expressivities of 17 objective-specification formalisms in RL. We place these formalisms in a preorder based on their expressive power, and present this preorder as a Hasse diagram. We find a variety of limitations for the different formalisms, and that no formalism is both dominantly expressive and straightforward to optimise with current techniques. For example, we prove that each of Regularised RL, Outer Nonlinear Markov Rewards, Reward Machines, Linear Temporal Logic, and Limit Average Rewards can express an objective that the others cannot. Our findings have implications for both policy optimisation and reward learning. Firstly, we identify expressivity limitations which are important to consider when specifying objectives in practice. Secondly, our results highlight the need for future research which adapts reward learning to work with a variety of formalisms, since many existing reward learning methods implicitly assume that desired objectives can be expressed with Markovian rewards. Our work contributes towards a more cohesive understanding of the costs and benefits of different RL objective-specification formalisms.

Variable speed limit (VSL) control is a promising traffic management strategy for enhancing safety and mobility. This work introduces MARVEL, a multi-agent reinforcement learning (MARL) framework for implementing large-scale VSL control on freeway corridors using only commonly available data. The agents learn through a reward structure that incorporates adaptability to traffic conditions, safety, and mobility; enabling coordination among the agents. The proposed framework scales to cover corridors with many gantries thanks to a parameter sharing among all VSL agents. The agents are trained in a microsimulation environment based on a short freeway stretch with 8 gantries spanning 7 miles and tested with 34 gantries spanning 17 miles of I-24 near Nashville, TN. MARVEL improves traffic safety by 63.4% compared to the no control scenario and enhances traffic mobility by 14.6% compared to a state-of-the-practice algorithm that has been deployed on I-24. An explainability analysis is undertaken to explore the learned policy under different traffic conditions and the results provide insights into the decision-making process of agents. Finally, we test the policy learned from the simulation-based experiments on real input data from I-24 to illustrate the potential deployment capability of the learned policy.

Synthesizing complex whole-body manipulation behaviors has fundamental challenges due to the rapidly growing combinatorics inherent to contact interaction planning. While model-based methods have shown promising results in solving long-horizon manipulation tasks, they often work under strict assumptions, such as known model parameters, oracular observation of the environment state, and simplified dynamics, resulting in plans that cannot easily transfer to hardware. Learning-based approaches, such as imitation learning (IL) and reinforcement learning (RL), have been shown to be robust when operating over in-distribution states; however, they need heavy human supervision. Specifically, model-free RL requires a tedious reward-shaping process. IL methods, on the other hand, rely on human demonstrations that involve advanced teleoperation methods. In this work, we propose a plan-guided reinforcement learning (PGRL) framework to combine the advantages of model-based planning and reinforcement learning. Our method requires minimal human supervision because it relies on plans generated by model-based planners to guide the exploration in RL. In exchange, RL derives a more robust policy thanks to domain randomization. We test this approach on a whole-body manipulation task on Punyo, an upper-body humanoid robot with compliant, air-filled arm coverings, to pivot and lift a large box. Our preliminary results indicate that the proposed methodology is promising to address challenges that remain difficult for either model- or learning-based strategies alone.

In inverse reinforcement learning (IRL), the central objective is to infer underlying reward functions from observed expert behaviors in a way that not only explains the given data but also generalizes to unseen scenarios. This ensures robustness against reward ambiguity where multiple reward functions can equally explain the same expert behaviors. While significant efforts have been made in addressing this issue, current methods often face challenges with high-dimensional problems and lack a geometric foundation. This paper harnesses the optimal transport (OT) theory to provide a fresh perspective on these challenges. By utilizing the Wasserstein distance from OT, we establish a geometric framework that allows for quantifying reward ambiguity and identifying a central representation or centroid of reward functions. These insights pave the way for robust IRL methodologies anchored in geometric interpretations, offering a structured approach to tackle reward ambiguity in high-dimensional settings.

Policy gradient methods have recently been shown to enjoy global convergence at a $\Theta(1/t)$ rate in the non-regularized tabular softmax setting. Accordingly, one important research question is whether this convergence rate can be further improved, with only first-order updates. In this paper, we answer the above question from the perspective of momentum by adapting the celebrated Nesterov's accelerated gradient (NAG) method to reinforcement learning (RL), termed \textit{Accelerated Policy Gradient} (APG). To demonstrate the potential of APG in achieving faster global convergence, we formally show that with the true gradient, APG with softmax policy parametrization converges to an optimal policy at a $\tilde{O}(1/t^2)$ rate. To the best of our knowledge, this is the first characterization of the global convergence rate of NAG in the context of RL. Notably, our analysis relies on one interesting finding: Regardless of the initialization, APG could end up reaching a locally nearly-concave regime, where APG could benefit significantly from the momentum, within finite iterations. By means of numerical validation, we confirm that APG exhibits $\tilde{O}(1/t^2)$ rate as well as show that APG could significantly improve the convergence behavior over the standard policy gradient.