Large transformer models powered by diverse data and model scale have dominated natural language modeling and computer vision and pushed the frontier of multiple AI areas. In reinforcement learning (RL), despite many efforts into transformer-based policies, a key limitation, however, is that current transformer-based policies cannot learn by directly combining information from multiple sub-optimal trials. In this work, we address this issue using recently proposed chain of hindsight to relabel experience, where we train a transformer on a sequence of trajectory experience ascending sorted according to their total rewards. Our method consists of relabelling target return of each trajectory to the maximum total reward among in sequence of trajectories and training an autoregressive model to predict actions conditioning on past states, actions, rewards, target returns, and task completion tokens, the resulting model, Agentic Transformer (AT), can learn to improve upon itself both at training and test time. As we show on D4RL and ExoRL benchmarks, to the best our knowledge, this is the first time that a simple transformer-based model performs competitively with both temporal-difference and imitation-learning-based approaches, even from sub-optimal data. Our Agentic Transformer also shows a promising scaling trend that bigger models consistently improve results.

We provide a new non-asymptotic analysis of distributed temporal difference learning with linear function approximation. Our approach relies on ``one-shot averaging,'' where $N$ agents run identical local copies of the TD(0) method and average the outcomes only once at the very end. We demonstrate a version of the linear time speedup phenomenon, where the convergence time of the distributed process is a factor of $N$ faster than the convergence time of TD(0). This is the first result proving benefits from parallelism for temporal difference methods.

Training visual reinforcement learning (RL) models in offline datasets is challenging due to overfitting issues in representation learning and overestimation problems in value function. In this paper, we propose a transfer learning method called Collaborative World Models (CoWorld) to improve the performance of visual RL under offline conditions. The core idea is to use an easy-to-interact, off-the-shelf simulator to train an auxiliary RL model as the online "test bed" for the offline policy learned in the target domain, which provides a flexible constraint for the value function -- Intuitively, we want to mitigate the overestimation problem of value functions outside the offline data distribution without impeding the exploration of actions with potential advantages. Specifically, CoWorld performs domain-collaborative representation learning to bridge the gap between online and offline hidden state distributions. Furthermore, it performs domain-collaborative behavior learning that enables the source RL agent to provide target-aware value estimation, allowing for effective offline policy regularization. Experiments show that CoWorld significantly outperforms existing methods in offline visual control tasks in DeepMind Control and Meta-World.

Games have been the perfect test-beds for artificial intelligence research for the characteristics that widely exist in real-world scenarios. Learning and optimisation, decision making in dynamic and uncertain environments, game theory, planning and scheduling, design and education are common research areas shared between games and real-world problems. Numerous open-source games or game-based environments have been implemented for studying artificial intelligence. In addition to single- or multi-player, collaborative or adversarial games, there has also been growing interest in implementing platforms for creative design in recent years. Those platforms provide ideal benchmarks for exploring and comparing artificial intelligence ideas and techniques. This paper reviews the game-based platforms for artificial intelligence research, discusses the research trend induced by the evolution of those platforms, and gives an outlook.

Most reinforcement learning algorithms treat the context under which they operate as a stationary, isolated and undisturbed environment. However, in the real world, the environment is constantly changing due to a variety of external influences. To address this problem, we study Markov Decision Processes (MDP) under the influence of an external temporal process. We formalize this notion and discuss conditions under which the problem becomes tractable with suitable solutions. We propose a policy iteration algorithm to solve this problem and theoretically analyze its performance.

Market making (MM) is an important research topic in quantitative finance, the agent needs to continuously optimize ask and bid quotes to provide liquidity and make profits. The limit order book (LOB) contains information on all active limit orders, which is an essential basis for decision-making. The modeling of evolving, high-dimensional and low signal-to-noise ratio LOB data is a critical challenge. Traditional MM strategy relied on strong assumptions such as price process, order arrival process, etc. Previous reinforcement learning (RL) works handcrafted market features, which is insufficient to represent the market. This paper proposes a RL agent for market making with LOB data. We leverage a neural network with convolutional filters and attention mechanism (Attn-LOB) for feature extraction from LOB. We design a new continuous action space and a hybrid reward function for the MM task. Finally, we conduct comprehensive experiments on latency and interpretability, showing that our agent has good applicability.

We hypothesize that taking advantage of prior knowledge about the underlying Figure 1: Workflow diagram for an existing self-paced RL task structure can improve the effectiveness approach, and two methods that we propose: intermediate of self-paced RL. We develop a self-paced RL self-paced RL and reward-machine-guided, self-paced RL. algorithm guided by reward machines, i.e., a type of finite-state machine that encodes the underlying task structure. The algorithm integrates reward machines in 1) the update of the policy and value et al. [2017] focus on automating the process of curriculum functions obtained by any RL algorithm of choice, generation. Klink et al. [2020a] adopt self-paced learning and 2) the update of the automated curriculum that [Kumar et al., 2010] in RL by developing an algorithm that generates context distributions. Our empirical results creates a sequence of probability distributions over contexts evidence that the proposed algorithm achieves [Hallak et al., 2015]. The dynamics, the reward function, optimal behavior reliably even in cases in which and the initial state distribution of an environment change existing baselines cannot make any meaningful with respect to the context. Given a target context distribution, progress. It also decreases the curriculum length a self-paced RL algorithm iteratively generates context and reduces the variance in the curriculum generation distributions that maximizes the expected discounted return, process by up to one-fourth and four orders of regularized by the Kullback-Leibler (KL) divergence from magnitude, respectively.

Counterfactual Risk Minimization (CRM) is a framework for dealing with the logged bandit feedback problem, where the goal is to improve a logging policy using offline data. In this paper, we explore the case where it is possible to deploy learned policies multiple times and acquire new data. We extend the CRM principle and its theory to this scenario, which we call "Sequential Counterfactual Risk Minimization (SCRM)." We introduce a novel counterfactual estimator and identify conditions that can improve the performance of CRM in terms of excess risk and regret rates, by using an analysis similar to restart strategies in accelerated optimization methods. We also provide an empirical evaluation of our method in both discrete and continuous action settings, and demonstrate the benefits of multiple deployments of CRM.

Stochastic gradient methods are an essential ingredient for solving various optimization problems, with a wide range of applications in various fields such as machine learning [Goodfellow et al., 2014, 2016], empirical risk minimization problems [Van der Vaart, 2000], and reinforcement learning [Sutton and Barto, 2018, Schulman et al., 2015, Mnih et al., 2015]. Various stochastic gradient descent methods (SGD) and their accelerated versions [Nesterov, 1983, Ghadimi and Lan, 2013] have been extensively studied under different statistical frameworks [Dieuleveut et al., 2017, Vaswani et al., 2019a]. The standard assumption for stochastic optimization algorithms is to consider independent and identically distributed noise variables. However, the growing usage of stochastic optimization methods in reinforcement learning [Bhandari et al., 2018, Srikant and Ying, 2019, Durmus et al., 2021] and distributed optimization [Lopes and Sayed, 2007, Dimakis et al., 2010, Mao et al., 2020] has led to increased interest in problems with Markovian noise. Despite this, existing theoretical works that consider Markov noise have significant limitations, and their analysis often results in suboptimal finite-time error bounds. Our research aims to fill the gap in the existing literature on the first-order Markovian setting. By focusing on uniformly geometrically ergodic Markov chains, we obtain finite-time complexity bounds for achieving ε-accurate solutions that scale linearly with the mixing time of the underlying Markov chain. Our approach is based on careful applications of randomized batch size schemes and provides a unified view on both non-convex and strongly convex minimization problems, as well as variational inequalities.

Developing learning-based methods for navigation of aerial robots is an intensive data-driven process that requires highly parallelized simulation. The full utilization of such simulators is hindered by the lack of parallelized high-level control methods that imitate the real-world robot interface. Responding to this need, we develop the Aerial Gym simulator that can simulate millions of multirotor vehicles parallelly with nonlinear geometric controllers for the Special Euclidean Group SE(3) for attitude, velocity and position tracking. We also develop functionalities for managing a large number of obstacles in the environment, enabling rapid randomization for learning of navigation tasks. In addition, we also provide sample environments having robots with simulated cameras capable of capturing RGB, depth, segmentation and optical flow data in obstacle-rich environments. This simulator is a step towards developing a - currently missing - highly parallelized aerial robot simulation with geometric controllers at a large scale, while also providing a customizable obstacle randomization functionality for navigation tasks. We provide training scripts with compatible reinforcement learning frameworks to navigate the robot to a goal setpoint based on attitude and velocity command interfaces. Finally, we open source the simulator and aim to develop it further to speed up rendering using alternate kernel-based frameworks in order to parallelize ray-casting for depth images thus supporting a larger number of robots.