It is well known that classical reinforcement learning, understood as learning from external rewards, has severe limitations. While it has been posited that reward is "enough" to learn any behavior [1], agents interacting with the real world often have only access to sparse rewards. Many approaches have been proposed to overcome the sparse reward limitation, endowing agents with additional signals to be optimized along with the rewards. These include minimizing surprise by refining predictions [2-7], novelty seeking by visiting states with low visit counts [8-10], generating actions that leads to predictable transitions (empowerment) [11-13], or seeking pure state entropy [14] and related forms of pure exploration objectives [3, 15-19], to name a few. A popular choice for augmenting the reward signal -the one that we focus on in this paper-is with entropy regularization [20-28]. The idea is that the agent will be driven, all else equal, to visit states and taking actions that make the agent act as random as possible (pure entropy regularization, e.g., [25]) or penalize the agent for having a policy very different from a default policy (KL regularization, e.g., [20]). Using this type of regularization can lead to better exploration [14], more variable and realistic behaviors [29], more efficient learning [25, 30] and more robust solutions [21] against noise and adversarial attacks [19] than classical reinforcement learning algorithms. While the above approaches use entropy as a regularizer to the optimization reward problem, the specific type of entropy regularizer varies widely across studies, and as a result the approaches and the solutions are hectic. For instance, some use pure action entropy regularization [24-26, 31], others employ purely state entropy [14], others take advantage of KL action regularization [23, 28, 32], and yet others combine action and state pure entropy in balanced [22, 33] or arbitrary ways [29].

The bandit paradigm provides a unified modeling framework for problems that require decision-making under uncertainty. Because many business metrics can be viewed as rewards (a.k.a. utilities) that result from actions, bandit algorithms have seen a large and growing interest from industrial applications, such as search, recommendation and advertising. Indeed, with the bandit lens comes the promise of direct optimisation for the metrics we care about. Nevertheless, the road to successfully applying bandits in production is not an easy one. Even when the action space and rewards are well-defined, practitioners still need to make decisions regarding multi-arm or contextual approaches, on- or off-policy setups, delayed or immediate feedback, myopic or long-term optimisation, etc. To make matters worse, industrial platforms typically give rise to large action spaces in which existing approaches tend to break down. The research literature on these topics is broad and vast, but this can overwhelm practitioners, whose primary aim is to solve practical problems, and therefore need to decide on a specific instantiation or approach for each project. This tutorial will take a step towards filling that gap between the theory and practice of bandits. Our goal is to present a unified overview of the field and its existing terminology, concepts and algorithms -- with a focus on problems relevant to industry. We hope our industrial perspective will help future practitioners who wish to leverage the bandit paradigm for their application.

RL is increasingly being used to control robotic systems that interact closely with humans. This interaction raises the problem of safe RL: how to ensure that a RL-controlled robotic system never, for instance, injures a human. This problem is especially challenging in rich, realistic settings where it is not even possible to clearly write down a reward function which incorporates these outcomes. In these circumstances, perhaps the only viable approach is based on IRL, which infers rewards from human demonstrations. However, IRL is massively underdetermined as many different rewards can lead to the same optimal policies; we show that this makes it difficult to distinguish catastrophic outcomes (such as injuring a human) from merely undesirable outcomes. Our key insight is that humans do display different behaviour when catastrophic outcomes are possible: they become much more careful. We incorporate carefulness signals into IRL, and find that they do indeed allow IRL to disambiguate undesirable from catastrophic outcomes, which is critical to ensuring safety in future real-world human-robot interactions.

Multiagent reinforcement learning (MARL) has benefited significantly from population-based and game-theoretic training regimes. One approach, Policy-Space Response Oracles (PSRO), employs standard reinforcement learning to compute response policies via approximate best responses and combines them via meta-strategy selection. We augment PSRO by adding a novel search procedure with generative sampling of world states, and introduce two new meta-strategy solvers based on the Nash bargaining solution. We evaluate PSRO's ability to compute approximate Nash equilibrium, and its performance in two negotiation games: Colored Trails, and Deal or No Deal. We conduct behavioral studies where human participants negotiate with our agents ($N = 346$). We find that search with generative modeling finds stronger policies during both training time and test time, enables online Bayesian co-player prediction, and can produce agents that achieve comparable social welfare negotiating with humans as humans trading among themselves.

We study the finite-horizon offline reinforcement learning (RL) problem. Since actions at any state can affect next-state distributions, the related distributional shift challenges can make this problem far more statistically complex than offline policy learning for a finite sequence of stochastic contextual bandit environments. We formalize this insight by showing that the statistical hardness of offline RL instances can be measured by estimating the size of actions' impact on next-state distributions. Furthermore, this estimated impact allows us to propagate just enough value function uncertainty from future steps to avoid model exploitation, enabling us to develop algorithms that improve upon traditional pessimistic approaches for offline RL on statistically simple instances. Our approach is supported by theory and simulations.

Modern reinforcement learning (RL) often faces an enormous state-action space. Existing analytical results are typically for settings with a small number of state-actions, or simple models such as linearly modeled Q-functions. To derive statistically efficient RL policies handling large state-action spaces, with more general Q-functions, some recent works have considered nonlinear function approximation using kernel ridge regression. In this work, we derive sample complexities for kernel based Q-learning when a generative model exists. We propose a nonparametric Q-learning algorithm which finds an $\epsilon$-optimal policy in an arbitrarily large scale discounted MDP. The sample complexity of the proposed algorithm is order optimal with respect to $\epsilon$ and the complexity of the kernel (in terms of its information gain). To the best of our knowledge, this is the first result showing a finite sample complexity under such a general model.

While recent advances in artificial intelligence have achieved human-level performance in environments like Starcraft and Go, many physical reasoning tasks remain challenging for modern algorithms. To date, few algorithms have been evaluated on physical tasks that involve manipulating objects when movable obstacles are present and when tools must be used to perform the manipulation. To promote research on such tasks, we introduce PushWorld, an environment with simplistic physics that requires manipulation planning with both movable obstacles and tools. We provide a benchmark of more than 200 PushWorld puzzles in PDDL and in an OpenAI Gym environment. We evaluate state-of-the-art classical planning and reinforcement learning algorithms on this benchmark, and we find that these baseline results are below human-level performance. We then provide a new classical planning heuristic that solves the most puzzles among the baselines, and although it is 40 times faster than the best baseline planner, it remains below human-level performance.

Machine learning (ML) tasks are one of the major workloads in today's edge computing networks. Existing edge-cloud schedulers allocate the requested amounts of resources to each task, falling short of best utilizing the limited edge resources for ML tasks. This paper proposes TapFinger, a distributed scheduler for edge clusters that minimizes the total completion time of ML tasks through co-optimizing task placement and fine-grained multi-resource allocation. To learn the tasks' uncertain resource sensitivity and enable distributed scheduling, we adopt multi-agent reinforcement learning (MARL) and propose several techniques to make it efficient, including a heterogeneous graph attention network as the MARL backbone, a tailored task selection phase in the actor network, and the integration of Bayes' theorem and masking schemes. We first implement a single-task scheduling version, which schedules at most one task each time. Then we generalize to the multi-task scheduling case, in which a sequence of tasks is scheduled simultaneously. Our design can mitigate the expanded decision space and yield fast convergence to optimal scheduling solutions. Extensive experiments using synthetic and test-bed ML task traces show that TapFinger can achieve up to 54.9% reduction in the average task completion time and improve resource efficiency as compared to state-of-the-art schedulers.

Reasoning in a complex and ambiguous environment is a key goal for Reinforcement Learning (RL) agents. While some sophisticated RL agents can successfully solve difficult tasks, they require a large amount of training data and often struggle to generalize to new unseen environments and new tasks. On the other hand, Large Scale Language Models (LSLMs) have exhibited strong reasoning ability and the ability to to adapt to new tasks through in-context learning. However, LSLMs do not inherently have the ability to interrogate or intervene on the environment. In this work, we investigate how to combine these complementary abilities in a single system consisting of three parts: a Planner, an Actor, and a Reporter. The Planner is a pre-trained language model that can issue commands to a simple embodied agent (the Actor), while the Reporter communicates with the Planner to inform its next command. We present a set of tasks that require reasoning, test this system's ability to generalize zero-shot and investigate failure cases, and demonstrate how components of this system can be trained with reinforcement-learning to improve performance.

Model-based next state prediction and state value prediction are slow to converge. To address these challenges, we do the following: i) Instead of a neural network, we do model-based planning using a parallel memory retrieval system (which we term the slow mechanism); ii) Instead of learning state values, we guide the agent's actions using goal-directed exploration, by using a neural network to choose the next action given the current state and the goal state (which we term the fast mechanism). The goal-directed exploration is trained online using hippocampal replay of visited states and future imagined states every single time step, leading to fast and efficient training. Empirical studies show that our proposed method has a 92% solve rate across 100 episodes in a dynamically changing grid world, significantly outperforming state-of-the-art actor critic mechanisms such as PPO (54%), TRPO (50%) and A2C (24%). Ablation studies demonstrate that both mechanisms are crucial. We posit that the future of Reinforcement Learning (RL) will be to model goals and sub-goals for various tasks, and plan it out in a goal-directed memory-based approach.