Reinforcement Learning (RL) has gained substantial attention across diverse application domains and theoretical investigations. Existing literature on RL theory largely focuses on risk-neutral settings where the decision-maker learns to maximize the expected cumulative reward. However, in practical scenarios such as portfolio management and e-commerce recommendations, decision-makers often persist in heterogeneous risk preferences subject to outcome uncertainties, which can not be well-captured by the risk-neural framework. Incorporating these preferences can be approached through utility theory, yet the development of risk-sensitive RL under general utility functions remains an open question for theoretical exploration. In this paper, we consider a scenario where the decision-maker seeks to optimize a general utility function of the cumulative reward in the framework of a Markov decision process (MDP). To facilitate the Dynamic Programming Principle and Bellman equation, we enlarge the state space with an additional dimension that accounts for the cumulative reward. We propose a discretized approximation scheme to the MDP under enlarged state space, which is tractable and key for algorithmic design. We then propose a modified value iteration algorithm that employs an epsilon-covering over the space of cumulative reward. When a simulator is accessible, our algorithm efficiently learns a near-optimal policy with guaranteed sample complexity. In the absence of a simulator, our algorithm, designed with an upper-confidence-bound exploration approach, identifies a near-optimal policy while ensuring a guaranteed regret bound. For both algorithms, we match the theoretical lower bounds for the risk-neutral setting.

The gameplay of strategic board games such as chess, Go and Hex is often characterized by combinatorial, relational structures--capturing distinct interactions and non-local patterns--and not just images. A key feature of CNNs is their relational inductive bias towards locality and translational invariance. Hence, we investigate the crucial question: Can GNNs, with their ability to encode complex connections, replace CNNs in self-play reinforcement learning? To this end, we do a comparison with Hex--an abstract yet strategically rich board game--serving as our experimental platform. Our findings reveal that GNNs excel at dealing with long range dependency situations in game states and are less prone to overfitting, but also showing a reduced proficiency in discerning local patterns. This suggests a potential paradigm shift, signaling the use of game-specific structures to reshape self-play reinforcement learning. In 2016, AlphaGo (Silver et al., 2016) became the first AI to beat professional Go players by combining Monte-Carlo tree search (MCTS) with neural network policy and value approximation and selfplay training. Its approach has since been transferred to various other board games: AlphaZero (Silver et al., 2017) achieved superhuman strength in Chess and Shogi while Mohex-3HNN (Gao et al., 2018) has become one of the strongest AI for Hex. Behind these accomplishments lies a crucial observation: Despite the diversity of the board games, all these programs use convolutional neural networks (CNN) as the foundational architecture to approximate policy and value functions. Figure 1: Non-local dependencies exist in many prominent board games. If A is blue, playing at B is a wasted move. These biases align harmoniously for most positions in the game of Go where spatially local patterns dominate.

A popular perspective in Reinforcement learning (RL) casts the problem as probabilistic inference on a graphical model of the Markov decision process (MDP). The core object of study is the probability of each state-action pair being visited under the optimal policy. Previous approaches to approximate this quantity can be arbitrarily poor, leading to algorithms that do not implement genuine statistical inference and consequently do not perform well in challenging problems. In this work, we undertake a rigorous Bayesian treatment of the posterior probability of state-action optimality and clarify how it flows through the MDP. We first reveal that this quantity can indeed be used to generate a policy that explores efficiently, as measured by regret. Unfortunately, computing it is intractable, so we derive a new variational Bayesian approximation yielding a tractable convex optimization problem and establish that the resulting policy also explores efficiently. We call our approach VAPOR and show that it has strong connections to Thompson sampling, K-learning, and maximum entropy exploration. We conclude with some experiments demonstrating the performance advantage of a deep RL version of VAPOR.

Completely Automated Public Turing test to tell Computers and Humans Apart, short for CAPTCHA, is an essential and relatively easy way to defend against malicious attacks implemented by bots. The security and usability trade-off limits the use of massive geometric transformations to interfere deep model recognition and deep models even outperformed humans in complex CAPTCHAs. The discovery of adversarial examples provides an ideal solution to the security and usability trade-off by integrating adversarial examples and CAPTCHAs to generate adversarial CAPTCHAs that can fool the deep models. In this paper, we extend the definition of adversarial CAPTCHAs and propose a classification method for adversarial CAPTCHAs. Then we systematically review some commonly used methods to generate adversarial examples and methods that are successfully used to generate adversarial CAPTCHAs. Also, we analyze some defense methods that can be used to defend adversarial CAPTCHAs, indicating potential threats to adversarial CAPTCHAs. Finally, we discuss some possible future research directions for adversarial CAPTCHAs at the end of this paper.

Portfolio management (PM) is a fundamental financial trading task, which explores the optimal periodical reallocation of capitals into different stocks to pursue long-term profits. Reinforcement learning (RL) has recently shown its potential to train profitable agents for PM through interacting with financial markets. However, existing work mostly focuses on fixed stock pools, which is inconsistent with investors' practical demand. Specifically, the target stock pool of different investors varies dramatically due to their discrepancy on market states and individual investors may temporally adjust stocks they desire to trade (e.g., adding one popular stocks), which lead to customizable stock pools (CSPs). Existing RL methods require to retrain RL agents even with a tiny change of the stock pool, which leads to high computational cost and unstable performance. To tackle this challenge, we propose EarnMore, a rEinforcement leARNing framework with Maskable stOck REpresentation to handle PM with CSPs through one-shot training in a global stock pool (GSP). Specifically, we first introduce a mechanism to mask out the representation of the stocks outside the target pool. Second, we learn meaningful stock representations through a self-supervised masking and reconstruction process. Third, a re-weighting mechanism is designed to make the portfolio concentrate on favorable stocks and neglect the stocks outside the target pool. Through extensive experiments on 8 subset stock pools of the US stock market, we demonstrate that EarnMore significantly outperforms 14 state-of-the-art baselines in terms of 6 popular financial metrics with over 40% improvement on profit.

Learning-based methods, particularly Reinforcement Learning (RL), hold great promise for streamlining deployment, enhancing performance, and achieving generalization in the control of autonomous multirotor aerial vehicles. Deep RL has been able to control complex systems with impressive fidelity and agility in simulation but the simulation-to-reality transfer often brings a hard-to-bridge reality gap. Moreover, RL is commonly plagued by prohibitively long training times. In this work, we propose a novel asymmetric actor-critic-based architecture coupled with a highly reliable RL-based training paradigm for end-to-end quadrotor control. We show how curriculum learning and a highly optimized simulator enhance sample complexity and lead to fast training times. To precisely discuss the challenges related to low-level/end-to-end multirotor control, we also introduce a taxonomy that classifies the existing levels of control abstractions as well as non-linearities and domain parameters. Our framework enables Simulation-to-Reality (Sim2Real) transfer for direct RPM control after only 18 seconds of training on a consumer-grade laptop as well as its deployment on microcontrollers to control a multirotor under real-time guarantees. Finally, our solution exhibits competitive performance in trajectory tracking, as demonstrated through various experimental comparisons with existing state-of-the-art control solutions using a real Crazyflie nano quadrotor. We open source the code including a very fast multirotor dynamics simulator that can simulate about 5 months of flight per second on a laptop GPU. The fast training times and deployment to a cheap, off-the-shelf quadrotor lower the barriers to entry and help democratize the research and development of these systems.

Although reinforcement learning methods offer a powerful framework for automatic skill acquisition, for practical learning-based control problems in domains such as robotics, imitation learning often provides a more convenient and accessible alternative. In particular, an interactive imitation learning method such as DAgger, which queries a near-optimal expert to intervene online to collect correction data for addressing the distributional shift challenges that afflict na\"ive behavioral cloning, can enjoy good performance both in theory and practice without requiring manually specified reward functions and other components of full reinforcement learning methods. In this paper, we explore how off-policy reinforcement learning can enable improved performance under assumptions that are similar but potentially even more practical than those of interactive imitation learning. Our proposed method uses reinforcement learning with user intervention signals themselves as rewards. This relaxes the assumption that intervening experts in interactive imitation learning should be near-optimal and enables the algorithm to learn behaviors that improve over the potential suboptimal human expert. We also provide a unified framework to analyze our RL method and DAgger; for which we present the asymptotic analysis of the suboptimal gap for both methods as well as the non-asymptotic sample complexity bound of our method. We then evaluate our method on challenging high-dimensional continuous control simulation benchmarks as well as real-world robotic vision-based manipulation tasks. The results show that it strongly outperforms DAgger-like approaches across the different tasks, especially when the intervening experts are suboptimal. Code and videos can be found on the project website: rlif-page.github.io

Same-Day Delivery (SDD) services aim to maximize the fulfillment of online orders while minimizing delivery delays but are beset by operational uncertainties such as those in order volumes and courier planning. Our work aims to enhance the operational efficiency of SDD by focusing on the ultra-fast Order Dispatching Problem (ODP), which involves matching and dispatching orders to couriers within a centralized warehouse setting, and completing the delivery within a strict timeline (e.g., within minutes). We introduce important extensions to ultra-fast ODP such as order batching and explicit courier assignments to provide a more realistic representation of dispatching operations and improve delivery efficiency. As a solution method, we primarily focus on NeurADP, a methodology that combines Approximate Dynamic Programming (ADP) and Deep Reinforcement Learning (DRL), and our work constitutes the first application of NeurADP outside of the ride-pool matching problem. NeurADP is particularly suitable for ultra-fast ODP as it addresses complex one-to-many matching and routing intricacies through a neural network-based VFA that captures high-dimensional problem dynamics without requiring manual feature engineering as in generic ADP methods. We test our proposed approach using four distinct realistic datasets tailored for ODP and compare the performance of NeurADP against myopic and DRL baselines by also making use of non-trivial bounds to assess the quality of the policies. Our numerical results indicate that the inclusion of order batching and courier queues enhances the efficiency of delivery operations and that NeurADP significantly outperforms other methods. Detailed sensitivity analysis with important parameters confirms the robustness of NeurADP under different scenarios, including variations in courier numbers, spatial setup, vehicle capacity, and permitted delay time.

Policies trained via reinforcement learning (RL) are often very complex even for simple tasks. In an episode with n time steps, a policy will make n decisions on actions to take, many of which may appear non-intuitive to the observer. Moreover, it is not clear which of these decisions directly contribute towards achieving the reward and how significant their contribution is. Given a trained policy, we propose a black-box method based on statistical covariance estimation that clusters the states of the environment and ranks each cluster according to the importance of decisions made in its states. We compare our measure against a previous statistical fault localization based ranking procedure.

In the domain of autonomous driving, the Learning from Demonstration (LfD) paradigm has exhibited notable efficacy in addressing sequential decision-making problems. However, consistently achieving safety in varying traffic contexts, especially in safety-critical scenarios, poses a significant challenge due to the long-tailed and unforeseen scenarios absent from offline datasets. In this paper, we introduce the saFety-aware strUctured Scenario representatION (FUSION), a pioneering methodology conceived to facilitate the learning of an adaptive end-to-end driving policy by leveraging structured scenario information. FUSION capitalizes on the causal relationships between decomposed reward, cost, state, and action space, constructing a framework for structured sequential reasoning under dynamic traffic environments. We conduct rigorous evaluations in two typical real-world settings of distribution shift in autonomous vehicles, demonstrating the good balance between safety cost and utility reward of FUSION compared to contemporary state-of-the-art safety-aware LfD baselines. Empirical evidence under diverse driving scenarios attests that FUSION significantly enhances the safety and generalizability of autonomous driving agents, even in the face of challenging and unseen environments. Furthermore, our ablation studies reveal noticeable improvements in the integration of causal representation into the safe offline RL problem.