Given a dataset on actions and resulting long-term rewards, a direct estimation approach fits value functions that minimize prediction error on the training data. Temporal difference learning (TD) methods instead fit value functions by minimizing the degree of temporal inconsistency between estimates made at successive time-steps. Focusing on finite state Markov chains, we provide a crisp asymptotic theory of the statistical advantages of this approach. First, we show that an intuitive inverse trajectory pooling coefficient completely characterizes the percent reduction in mean-squared error of value estimates. Depending on problem structure, the reduction could be enormous or nonexistent. Next, we prove that there can be dramatic improvements in estimates of the difference in value-to-go for two states: TD's errors are bounded in terms of a novel measure - the problem's trajectory crossing time - which can be much smaller than the problem's time horizon.

Deep reinforcement learning repeatedly succeeds in closed, well-defined domains such as games (Chess, Go, StarCraft). The next frontier is real-world scenarios, where setups are numerous and varied. For this, agents need to learn the underlying rules governing the environment, so as to robustly generalise to conditions that differ from those they were trained on. Model-based reinforcement learning algorithms, such as the highly successful MuZero, aim to accomplish this by learning a world model. However, leveraging a world model has not consistently shown greater generalisation capabilities compared to model-free alternatives. In this work, we propose improving the data efficiency and generalisation capabilities of MuZero by explicitly incorporating the symmetries of the environment in its world-model architecture. We prove that, so long as the neural networks used by MuZero are equivariant to a particular symmetry group acting on the environment, the entirety of MuZero's action-selection algorithm will also be equivariant to that group. We evaluate Equivariant MuZero on procedurally-generated MiniPacman and on Chaser from the ProcGen suite: training on a set of mazes, and then testing on unseen rotated versions, demonstrating the benefits of equivariance. Further, we verify that our performance improvements hold even when only some of the components of Equivariant MuZero obey strict equivariance, which highlights the robustness of our construction.

Training a team to complete a complex task via multi-agent reinforcement learning can be difficult due to challenges such as policy search in a large policy space, and non-stationarity caused by mutually adapting agents. To facilitate efficient learning of complex multi-agent tasks, we propose an approach which uses an expert-provided curriculum of simpler multi-agent sub-tasks. In each sub-task of the curriculum, a subset of the entire team is trained to acquire sub-task-specific policies. The sub-teams are then merged and transferred to the target task, where their policies are collectively fined tuned to solve the more complex target task. We present MEDoE, a flexible method which identifies situations in the target task where each agent can use its sub-task-specific skills, and uses this information to modulate hyperparameters for learning and exploration during the fine-tuning process. We compare MEDoE to multi-agent reinforcement learning baselines that train from scratch in the full task, and with na\"ive applications of standard multi-agent reinforcement learning techniques for fine-tuning. We show that MEDoE outperforms baselines which train from scratch or use na\"ive fine-tuning approaches, requiring significantly fewer total training timesteps to solve a range of complex teamwork tasks.

How is it that humans can solve complex planning tasks so efficiently despite limited cognitive resources? One reason is its ability to know how to use its limited computational resources to make clever choices. We postulate that people learn this ability from trial and error (metacognitive reinforcement learning). Here, we systematize models of the underlying learning mechanisms and enhance them with more sophisticated additional mechanisms. We fit the resulting 86 models to human data collected in previous experiments where different phenomena of metacognitive learning were demonstrated and performed Bayesian model selection. Our results suggest that a gradient ascent through the space of cognitive strategies can explain most of the observed qualitative phenomena, and is therefore a promising candidate for explaining the mechanism underlying metacognitive learning.

Mobile edge computing (MEC) is a promising paradigm to meet the quality of service (QoS) requirements of latency-sensitive IoT applications. However, attackers may eavesdrop on the offloading decisions to infer the edge server's (ES's) queue information and users' usage patterns, thereby incurring the pattern privacy (PP) issue. Therefore, we propose an offloading strategy which jointly minimizes the latency, ES's energy consumption, and task dropping rate, while preserving PP. Firstly, we formulate the dynamic computation offloading procedure as a Markov decision process (MDP). Next, we develop a Differential Privacy Deep Q-learning based Offloading (DP-DQO) algorithm to solve this problem while addressing the PP issue by injecting noise into the generated offloading decisions. This is achieved by modifying the deep Q-network (DQN) with a Function-output Gaussian process mechanism. We provide a theoretical privacy guarantee and a utility guarantee (learning error bound) for the DP-DQO algorithm and finally, conduct simulations to evaluate the performance of our proposed algorithm by comparing it with greedy and DQN-based algorithms.

We prove that optimistic-follow-the-regularized-leader (OFTRL), together with smooth value updates, finds an $O(T^{-1})$-approximate Nash equilibrium in $T$ iterations for two-player zero-sum Markov games with full information. This improves the $\tilde{O}(T^{-5/6})$ convergence rate recently shown in the paper Zhang et al (2022). The refined analysis hinges on two essential ingredients. First, the sum of the regrets of the two players, though not necessarily non-negative as in normal-form games, is approximately non-negative in Markov games. This property allows us to bound the second-order path lengths of the learning dynamics. Second, we prove a tighter algebraic inequality regarding the weights deployed by OFTRL that shaves an extra $\log T$ factor. This crucial improvement enables the inductive analysis that leads to the final $O(T^{-1})$ rate.

In recent years, reinforcement learning and its multi-agent analogue have achieved great success in solving various complex control problems. However, multi-agent reinforcement learning remains challenging both in its theoretical analysis and empirical design of algorithms, especially for large swarms of embodied robotic agents where a definitive toolchain remains part of active research. We use emerging state-of-the-art mean-field control techniques in order to convert many-agent swarm control into more classical single-agent control of distributions. This allows profiting from advances in single-agent reinforcement learning at the cost of assuming weak interaction between agents. However, the mean-field model is violated by the nature of real systems with embodied, physically colliding agents. Thus, we combine collision avoidance and learning of mean-field control into a unified framework for tractably designing intelligent robotic swarm behavior. On the theoretical side, we provide novel approximation guarantees for general mean-field control both in continuous spaces and with collision avoidance. On the practical side, we show that our approach outperforms multi-agent reinforcement learning and allows for decentralized open-loop application while avoiding collisions, both in simulation and real UAV swarms. Overall, we propose a framework for the design of swarm behavior that is both mathematically well-founded and practically useful, enabling the solution of otherwise intractable swarm problems.

Reinforcement learning generalizes bandit problems with additional difficulties on longer planning horizon and unknown transition kernel. We show that, under some mild assumptions, *any* slowly changing adversarial bandit algorithm enjoys optimal regret in adversarial bandits can achieve optimal (in the dependency of $T$) expected regret in infinite-horizon discounted MDPs, without the presence of Bellman backups. The slowly changing property required by our generalization is mild, which is also marked by the online Markov decision process literature. We also examine the applicability of our reduction to a well-known adversarial bandit algorithm, EXP3.

Meta Reinforcement Learning (Meta-RL) has seen substantial advancements recently. In particular, off-policy methods were developed to improve the data efficiency of Meta-RL techniques. \textit{Probabilistic embeddings for actor-critic RL} (PEARL) is a leading approach for multi-MDP adaptation problems. A major drawback of many existing Meta-RL methods, including PEARL, is that they do not explicitly consider the safety of the prior policy when it is exposed to a new task for the first time. Safety is essential for many real-world applications, including field robots and Autonomous Vehicles (AVs). In this paper, we develop the PEARL PLUS (PEARL$^+$) algorithm, which optimizes the policy for both prior (pre-adaptation) safety and posterior (after-adaptation) performance. Building on top of PEARL, our proposed PEARL$^+$ algorithm introduces a prior regularization term in the reward function and a new Q-network for recovering the state-action value under prior context assumptions, to improve the robustness to task distribution shift and safety of the trained network exposed to a new task for the first time. The performance of PEARL$^+$ is validated by solving three safety-critical problems related to robots and AVs, including two MuJoCo benchmark problems. From the simulation experiments, we show that safety of the prior policy is significantly improved and more robust to task distribution shift compared to PEARL.

Context information is in demand more than ever with the rapid increase in the number of context-aware Internet of Things applications developed worldwide. Research in context and context-awareness is being conducted to broaden its applicability in light of many practical and technical challenges. One of the challenges is improving performance when responding to a large number of context queries. Context Management Platforms that infer and deliver context to applications measure this problem using Quality of Service (QoS) parameters. Although caching is a proven way to improve QoS, transiency of context and features such as variability and heterogeneity of context queries pose an additional real-time cost management problem. This paper presents a critical survey of the state-of-the-art in adaptive data caching with the objective of developing a body of knowledge in cost- and performance-efficient adaptive caching strategies. We comprehensively survey a large number of research publications and evaluate, compare, and contrast different techniques, policies, approaches, and schemes in adaptive caching. Our critical analysis is motivated by the focus on adaptively caching context as a core research problem. A formal definition for adaptive context caching is then proposed, followed by identified features and requirements of a well-designed, objective optimal adaptive context caching strategy.