Reinforcement learning (RL, Sutton & Barto, 2018) is a general technique in sequential decision making that learns an optimal policy to maximize the average cumulative reward. Prior to adopting any policy in practice, it is crucial to know the impact of implementing such a policy. In many real domains such as healthcare (Murphy et al., 2001; Luedtke & van der Laan, 2017; Shi et al., 2020a), robotics (Andrychowicz et al., 2020) and autonomous driving (Sallab et al., 2017), it is costly, risky, unethical, or even infeasible to evaluate a policy's impact by directly running this policy. This motivates us to study the off-policy evaluation (OPE) problem that learns a target policy's value with pre-collected data generated by a different behavior policy. In many applications (e.g., mobile health studies), the number of observations is limited. Take the OhioT1DM dataset (Marling & Bunescu, 2018) as an example, only a few thousands observations are available (Shi et al., 2020b). In these cases, in addition to a point estimate on a target policy's value, it is crucial to construct a confidence interval (CI) that quantifies the uncertainty of the value estimates. This paper is concerned with the following question: is it possible to develop a robust and efficient off-policy value estimator, and provide rigorous uncertainty quantification under practically feasible conditions? We will give an affirmative answer to this question.

Reinforcement learning lies at the intersection of several challenges. Many applications of interest involve extremely large state spaces, requiring function approximation to enable tractable computation. In addition, the learner has only a single stream of experience with which to evaluate a large number of possible courses of action, necessitating algorithms which can learn off-policy. However, the combination of off-policy learning with function approximation leads to divergence of temporal difference methods. Recent work into gradient-based temporal difference methods has promised a path to stability, but at the cost of expensive hyperparameter tuning. In parallel, progress in online learning has provided parameter-free methods that achieve minimax optimal guarantees up to logarithmic terms, but their application in reinforcement learning has yet to be explored. In this work, we combine these two lines of attack, deriving parameter-free, gradient-based temporal difference algorithms. Our algorithms run in linear time and achieve high-probability convergence guarantees matching those of GTD2 up to $\log$ factors. Our experiments demonstrate that our methods maintain high prediction performance relative to fully-tuned baselines, with no tuning whatsoever.

The performance of deep reinforcement learning methods prone to degenerate when applied to environments with non-stationary dynamics. In this paper, we utilize the latent context recurrent encoders motivated by recent Meta-RL materials, and propose the Latent Context-based Soft Actor Critic (LC-SAC) method to address aforementioned issues. By minimizing the contrastive prediction loss function, the learned context variables capture the information of the environment dynamics and the recent behavior of the agent. Then combined with the soft policy iteration paradigm, the LC-SAC method alternates between soft policy evaluation and soft policy improvement until it converges to the optimal policy. Experimental results show that the performance of LC-SAC is significantly better than the SAC algorithm on the MetaWorld ML1 tasks whose dynamics changes drasticly among different episodes, and is comparable to SAC on the continuous control benchmark task MuJoCo whose dynamics changes slowly or doesn't change between different episodes. In addition, we also conduct relevant experiments to determine the impact of different hyperparameter settings on the performance of the LC-SAC algorithm and give the reasonable suggestions of hyperparameter setting.

This course is all about the application of deep learning and neural networks to reinforcement learning. If you've taken my first reinforcement learning class, then you know that reinforcement learning is on the bleeding edge of what we can do with AI. Specifically, the combination of deep learning with reinforcement learning has led to AlphaGo beating a world champion in the strategy game Go, it has led to self-driving cars, and it has led to machines that can play video games at a superhuman level. Reinforcement learning has been around since the 70s but none of this has been possible until now. The world is changing at a very fast pace.

Markov Decision Processes (MDPs) play key mathematical models in Reinforcement Learning (RL). Despite its success in empirical performances [Haarnoja et al., 2018, Mnih et al., 2015, 2016, Silver et al., 2016], there are also many works providing insightful and solid theoretical understandings towards RL. The difficulty of solving an MDP mainly is due to the reward and transition probability, whose exact information is usually unknown to observers. To deal with the situations, one common approach resorts to offline methods, where the agent only has access to a given explorable dataset generated by given strategies. Many practical deep RL algorithms employ the offline method and achieve state-of-art success [Mnih et al., 2015, Lillicrap et al., 2015, Fujimoto et al., 2019]. In addition to empirical success, there are flourishing works on offline RL from a theoretical perspective. Some prior works [Chen and Jiang, 2019, Agarwal et al., 2020, Duan et al., 2021] have provided solid results on model-free offline methods, while some other works [Sidford et al., 2018, Xie et al., 2019, Yin and Wang, 2020, Yin et al., 2020] consider model-based approaches. However, Mannor et al. [2004] showed that model-based approaches can be quite sensitive to estimation errors by directly estimating the transition probability from an offline dataset.

Scene Rearrangement Planning (SRP) is an interior task proposed recently. The previous work defines the action space of this task with handcrafted coarse-grained actions that are inflexible to be used for transforming scene arrangement and intractable to be deployed in practice. Additionally, this new task lacks realistic indoor scene rearrangement data to feed popular data-hungry learning approaches and meet the needs of quantitative evaluation. To address these problems, we propose a fine-grained action definition for SRP and introduce a large-scale scene rearrangement dataset. We also propose a novel learning paradigm to efficiently train an agent through self-playing, without any prior knowledge. The agent trained via our paradigm achieves superior performance on the introduced dataset compared to the baseline agents. We provide a detailed analysis of the design of our approach in our experiments.

Building on the breakthrough of reinforcement learning, this paper introduces a unified framework of model-free reinforcement learning, CASA-B, Critic AS an Actor with Bandits Vote Algorithm. CASA-B is an actor-critic framework that estimates state-value, state-action-value and policy. An expectation-correct Doubly Robust Trace is introduced to learn state-value and state-action-value, whose convergence properties are guaranteed. We prove that CASA-B integrates a consistent path for the policy evaluation and the policy improvement. The policy evaluation is equivalent to a compensational policy improvement, which alleviates the function approximation error, and is also equivalent to an entropy-regularized policy improvement, which prevents the policy from collapsing to a suboptimal solution. Building on this design, we find the entropy of the behavior policies' and the target policy's are disentangled. Based on this observation, we propose a progressive closed-form entropy control mechanism, which explicitly controls the behavior policies' entropy to arbitrary range. Our experiments show that CASAB is super sample efficient and achieves State-Of-The-Art on Arcade Learning Environment. Our mean Human Normalized Score is 6456.63% and our median Human Normalized Score is 477.17%, under 200M training scale.

In recent years, quantitative investment methods combined with artificial intelligence have attracted more and more attention from investors and researchers. Existing related methods based on the supervised learning are not very suitable for learning problems with long-term goals and delayed rewards in real futures trading. In this paper, therefore, we model the price prediction problem as a Markov decision process (MDP), and optimize it by reinforcement learning with expert trajectory. In the proposed method, we employ more than 100 short-term alpha factors instead of price, volume and several technical factors in used existing methods to describe the states of MDP. Furthermore, unlike DQN (deep Q-learning) and BC (behavior cloning) in related methods, we introduce expert experience in training stage, and consider both the expert-environment interaction and the agent-environment interaction to design the temporal difference error so that the agents are more adaptable for inevitable noise in financial data. Experimental results evaluated on share price index futures in China, including IF (CSI 300) and IC (CSI 500), show that the advantages of the proposed method compared with three typical technical analysis and two deep leaning based methods.

Professional race-car drivers can execute extreme overtaking maneuvers. However, existing algorithms for autonomous overtaking either rely on simplified assumptions about the vehicle dynamics or try to solve expensive trajectory-optimization problems online. When the vehicle approaches its physical limits, existing model-based controllers struggle to handle highly nonlinear dynamics, and cannot leverage the large volume of data generated by simulation or real-world driving. To circumvent these limitations, we propose a new learning-based method to tackle the autonomous overtaking problem. We evaluate our approach in the popular car racing game Gran Turismo Sport, which is known for its detailed modeling of various cars and tracks. By leveraging curriculum learning, our approach leads to faster convergence as well as increased performance compared to vanilla reinforcement learning. As a result, the trained controller outperforms the built-in model-based game AI and achieves comparable overtaking performance with an experienced human driver.

Model-free deep reinforcement learning has achieved great success in many domains, such as video games, recommendation systems and robotic control tasks. In continuous control tasks, widely used policies with Gaussian distributions results in ineffective exploration of environments and limited performance of algorithms in many cases. In this paper, we propose a density-free off-policy algorithm, Generative Actor-Critic(GAC), using the push-forward model to increase the expressiveness of policies, which also includes an entropy-like technique, MMD-entropy regularizer, to balance the exploration and exploitation. Additionnally, we devise an adaptive mechanism to automatically scale this regularizer, which further improves the stability and robustness of GAC. The experiment results show that push-forward policies possess desirable features, such as multi-modality, which can improve the efficiency of exploration and asymptotic performance of algorithms obviously.