Using simulations between pairs of $\epsilon$-greedy q-learners with one-period memory, this article demonstrates that the potential function of the stochastic replicator dynamics (Foster and Young, 1990) allows it to predict the emergence of error-proof cooperative strategies from the underlying parameters of the repeated prisoner's dilemma. The observed cooperation rates between q-learners are related to the ratio between the kinetic energy exerted by the polar attractors of the replicator dynamics under the grim trigger strategy. The frontier separating the parameter space conducive to cooperation from the parameter space dominated by defection can be found by setting the kinetic energy ratio equal to a critical value, which is a function of the discount factor, $f(\delta) = \delta/(1-\delta)$, multiplied by a correction term to account for the effect of the algorithms' exploration probability. The gradient at the frontier increases with the distance between the game parameters and the hyperplane that characterizes the incentive compatibility constraint for cooperation under grim trigger. Building on literature from the neurosciences, which suggests that reinforcement learning is useful to understanding human behavior in risky environments, the article further explores the extent to which the frontier derived for q-learners also explains the emergence of cooperation between humans. Using metadata from laboratory experiments that analyze human choices in the infinitely repeated prisoner's dilemma, the cooperation rates between humans are compared to those observed between q-learners under similar conditions. The correlation coefficients between the cooperation rates observed for humans and those observed for q-learners are consistently above $0.8$. The frontier derived from the simulations between q-learners is also found to predict the emergence of cooperation between humans.

Motivated by personalized healthcare and other applications involving sensitive data, we study online exploration in reinforcement learning with differential privacy (DP) constraints. Existing work on this problem established that no-regret learning is possible under joint differential privacy (JDP) and local differential privacy (LDP) but did not provide an algorithm with optimal regret. We close this gap for the JDP case by designing an $\epsilon$-JDP algorithm with a regret of $\widetilde{O}(\sqrt{SAH^2T}+S^2AH^3/\epsilon)$ which matches the information-theoretic lower bound of non-private learning for all choices of $\epsilon> S^{1.5}A^{0.5} H^2/\sqrt{T}$. In the above, $S$, $A$ denote the number of states and actions, $H$ denotes the planning horizon, and $T$ is the number of steps. To the best of our knowledge, this is the first private RL algorithm that achieves \emph{privacy for free} asymptotically as $T\rightarrow \infty$. Our techniques -- which could be of independent interest -- include privately releasing Bernstein-type exploration bonuses and an improved method for releasing visitation statistics. The same techniques also imply a slightly improved regret bound for the LDP case.

We investigate the problem of autonomous racing among teams of cooperative agents that are subject to realistic racing rules. Our work extends previous research on hierarchical control in head-to-head autonomous racing by considering a generalized version of the problem while maintaining the two-level hierarchical control structure. A high-level tactical planner constructs a discrete game that encodes the complex rules using simplified dynamics to produce a sequence of target waypoints. The low-level path planner uses these waypoints as a reference trajectory and computes high-resolution control inputs by solving a simplified formulation of a racing game with a simplified representation of the realistic racing rules. We explore two approaches for the low-level path planner: training a multi-agent reinforcement learning (MARL) policy and solving a linear-quadratic Nash game (LQNG) approximation. We evaluate our controllers on simple and complex tracks against three baselines: an end-to-end MARL controller, a MARL controller tracking a fixed racing line, and an LQNG controller tracking a fixed racing line. Quantitative results show our hierarchical methods outperform the baselines in terms of race wins, overall team performance, and compliance with the rules. Qualitatively, we observe the hierarchical controllers mimic actions performed by expert human drivers such as coordinated overtaking, defending against multiple opponents, and long-term planning for delayed advantages.

Causal discovery is a major task with the utmost importance for machine learning since causal structures can enable models to go beyond pure correlation-based inference and significantly boost their performance. However, finding causal structures from data poses a significant challenge both in computational effort and accuracy, let alone its impossibility without interventions in general. In this paper, we develop a meta-reinforcement learning algorithm that performs causal discovery by learning to perform interventions such that it can construct an explicit causal graph. Apart from being useful for possible downstream applications, the estimated causal graph also provides an explanation for the data-generating process. In this article, we show that our algorithm estimates a good graph compared to the SOTA approaches, even in environments whose underlying causal structure is previously unseen. Further, we make an ablation study that shows how learning interventions contribute to the overall performance of our approach. We conclude that interventions indeed help boost the performance, efficiently yielding an accurate estimate of the causal structure of a possibly unseen environment.

Learning to collaborate is critical in Multi-Agent Reinforcement Learning (MARL). Previous works promote collaboration by maximizing the correlation of agents' behaviors, which is typically characterized by Mutual Information (MI) in different forms. However, we reveal sub-optimal collaborative behaviors also emerge with strong correlations, and simply maximizing the MI can, surprisingly, hinder the learning towards better collaboration. To address this issue, we propose a novel MARL framework, called Progressive Mutual Information Collaboration (PMIC), for more effective MI-driven collaboration. PMIC uses a new collaboration criterion measured by the MI between global states and joint actions. Based on this criterion, the key idea of PMIC is maximizing the MI associated with superior collaborative behaviors and minimizing the MI associated with inferior ones. The two MI objectives play complementary roles by facilitating better collaborations while avoiding falling into sub-optimal ones. Experiments on a wide range of MARL benchmarks show the superior performance of PMIC compared with other algorithms.

Robotic locomotion is often approached with the goal of maximizing robustness and reactivity by increasing motion control frequency. We challenge this intuitive notion by demonstrating robust and dynamic locomotion with a learned motion controller executing at as low as 8 Hz on a real ANYmal C quadruped. The robot is able to robustly and repeatably achieve a high heading velocity of 1.5 m/s, traverse uneven terrain, and resist unexpected external perturbations. We further present a comparative analysis of deep reinforcement learning (RL) based motion control policies trained and executed at frequencies ranging from 5 Hz to 200 Hz. We show that low-frequency policies are less sensitive to actuation latencies and variations in system dynamics. This is to the extent that a successful sim-to-real transfer can be performed even without any dynamics randomization or actuation modeling. We support this claim through a set of rigorous empirical evaluations. Moreover, to assist reproducibility, we provide the training and deployment code along with an extended analysis at https://ori-drs.github.io/lfmc/.

Large-scale finite element simulations of complex physical systems governed by partial differential equations (PDE) crucially depend on adaptive mesh refinement (AMR) to allocate computational budget to regions where higher resolution is required. Existing scalable AMR methods make heuristic refinement decisions based on instantaneous error estimation and thus do not aim for long-term optimality over an entire simulation. We propose a novel formulation of AMR as a Markov decision process and apply deep reinforcement learning (RL) to train refinement policies directly from simulation. AMR poses a new problem for RL as both the state dimension and available action set changes at every step, which we solve by proposing new policy architectures with differing generality and inductive bias. The model sizes of these policy architectures are independent of the mesh size and hence can be deployed on larger simulations than those used at train time. We demonstrate in comprehensive experiments on static function estimation and time-dependent equations that RL policies can be trained on problems without using ground truth solutions, are competitive with a widely-used error estimator, and generalize to larger, more complex, and unseen test problems.

We present a novel AI-assisted method for decomposing (segmenting) planar CAD (computer-aided design) models into well shaped rectangular blocks as a proof-of-principle of a general decomposition method applicable to complex 2D and 3D CAD models. The decomposed blocks are required for generating good quality meshes (tilings of quadrilaterals or hexahedra) suitable for numerical simulations of physical systems governed by conservation laws. The problem of hexahedral mesh generation of general CAD models has vexed researchers for over 3 decades and analysts often spend more than 50% of the design-analysis cycle time decomposing complex models into simpler parts meshable by existing techniques. Our method uses reinforcement learning to train an agent to perform a series of optimal cuts on the CAD model that result in a good quality block decomposition. We show that the agent quickly learns an effective strategy for picking the location and direction of the cuts and maximizing its rewards as opposed to making random cuts. This paper is the first successful demonstration of an agent autonomously learning how to perform this block decomposition task effectively thereby holding the promise of a viable method to automate this challenging process.

In this work, we ask for and answer what makes classical temporal-difference reinforcement learning with epsilon-greedy strategies cooperative. Cooperating in social dilemma situations is vital for animals, humans, and machines. While evolutionary theory revealed a range of mechanisms promoting cooperation, the conditions under which agents learn to cooperate are contested. Here, we demonstrate which and how individual elements of the multi-agent learning setting lead to cooperation. We use the iterated Prisoner's dilemma with one-period memory as a testbed. Each of the two learning agents learns a strategy that conditions the following action choices on both agents' action choices of the last round. We find that next to a high caring for future rewards, a low exploration rate, and a small learning rate, it is primarily intrinsic stochastic fluctuations of the reinforcement learning process which double the final rate of cooperation to up to 80%. Thus, inherent noise is not a necessary evil of the iterative learning process. It is a critical asset for the learning of cooperation. However, we also point out the trade-off between a high likelihood of cooperative behavior and achieving this in a reasonable amount of time. Our findings are relevant for purposefully designing cooperative algorithms and regulating undesired collusive effects.

Speaker diarization is a crucial task in many real-world applications, such as meeting transcription, call center monitoring, and broadcast news processing. The goal of speaker diarization is to partition an audio or video stream into homogeneous segments, each corresponding to a single speaker, without any prior knowledge of the speakers' identities [1, 2]. This task has traditionally been addressed using unsupervised clustering methods [3, 4, 5], but recent advances in deep learning have led to the development of more powerful embedding-based approaches [6, 7, 5]. Despite the recent progress, speaker diarization remains a challenging problem, particularly in real-time and online scenarios where new speakers may enter or leave the conversation at any time. In such cases, pre-trained models may not be sufficient, and the system must be able to adapt to new speakers on the fly [8, 9, 10]. As in the successful applications to other speech and language tasks [11], the reinforcement learning (RL) has emerged as a promising approach for developing next-generation speaker diarization systems that can learn online and adapt to changing circumstances. In this paper, we propose a novel RL framework for online speaker diarization that does not require prior registration or pretraining. Our approach combines embedding extraction, clustering, and resegmentation into a single online decision-making problem, where the agent receives feedback in the form of rewards or penalties for each segmentation decision. We demonstrate the effectiveness of our approach using a Q-learning-based diarization agent on a desktop app, and discuss practical considerations for implementing and deploying RL-based speaker diarization systems.