While deep reinforcement learning (RL) agents outperform humans on an increasing number of tasks, training them requires data equivalent to decades of human gameplay. Recent hierarchical RL methods have increased sample efficiency by incorporating information inherent to the structure of the decision problem but at the cost of having to discover or use human-annotated sub-goals that guide the learning process. We show that intentions of human players, i.e. the precursor of goal-oriented decisions, can be robustly predicted from eye gaze even for the long-horizon sparse rewards task of Montezuma's Revenge - one of the most challenging RL tasks in the Atari2600 game suite. We propose Int-HRL: Hierarchical RL with intention-based sub-goals that are inferred from human eye gaze. Our novel sub-goal extraction pipeline is fully automatic and replaces the need for manual sub-goal annotation by human experts. Our evaluations show that replacing hand-crafted sub-goals with automatically extracted intentions leads to a HRL agent that is significantly more sample efficient than previous methods.

Traffic scenarios in roundabouts pose substantial complexity for automated driving. Manually mapping all possible scenarios into a state space is labor-intensive and challenging. Deep reinforcement learning (DRL) with its ability to learn from interacting with the environment emerges as a promising solution for training such automated driving models. This study explores, employs, and implements various DRL algorithms, namely Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), and Trust Region Policy Optimization (TRPO) to instruct automated vehicles' driving through roundabouts. The driving state space, action space, and reward function are designed. The reward function considers safety, efficiency, comfort, and energy consumption to align with real-world requirements. All three tested DRL algorithms succeed in enabling automated vehicles to drive through the roundabout. To holistically evaluate the performance of these algorithms, this study establishes an evaluation methodology considering multiple indicators such as safety, efficiency, and comfort level. A method employing the Analytic Hierarchy Process is also developed to weigh these evaluation indicators. Experimental results on various testing scenarios reveal that the TRPO algorithm outperforms DDPG and PPO in terms of safety and efficiency, and PPO performs best in terms of comfort level. Lastly, to verify the model's adaptability and robustness regarding other driving scenarios, this study also deploys the model trained by TRPO to a range of different testing scenarios, e.g., highway driving and merging. Experimental results demonstrate that the TRPO model trained on only roundabout driving scenarios exhibits a certain degree of proficiency in highway driving and merging scenarios. This study provides a foundation for the application of automated driving with DRL in real traffic environments.

In a broad class of reinforcement learning applications, stochastic rewards have heavy-tailed distributions, which lead to infinite second-order moments for stochastic (semi)gradients in policy evaluation and direct policy optimization. In such instances, the existing RL methods may fail miserably due to frequent statistical outliers. In this work, we establish that temporal difference (TD) learning with a dynamic gradient clipping mechanism, and correspondingly operated natural actor-critic (NAC), can be provably robustified against heavy-tailed reward distributions. It is shown in the framework of linear function approximation that a favorable tradeoff between bias and variability of the stochastic gradients can be achieved with this dynamic gradient clipping mechanism. In particular, we prove that robust versions of TD learning achieve sample complexities of order $\mathcal{O}(\varepsilon^{-\frac{1}{p}})$ and $\mathcal{O}(\varepsilon^{-1-\frac{1}{p}})$ with and without the full-rank assumption on the feature matrix, respectively, under heavy-tailed rewards with finite moments of order $(1+p)$ for some $p\in(0,1]$, both in expectation and with high probability. We show that a robust variant of NAC based on Robust TD learning achieves $\tilde{\mathcal{O}}(\varepsilon^{-4-\frac{2}{p}})$ sample complexity. We corroborate our theoretical results with numerical experiments.

The evolution of cooperation has been investigated intensely in various disciplines, such as biology, economics, computer science, physics and psychology. There are two important dimensions, among many (Bowles and Gintis, 2011; Lehmann and Keller, 2006; Nowak, 2006), that have been shown to affect the evolution of cooperation: the interaction structure, i.e., who interacts with whom (Santos et al., 2006), and the mode of cognition, i.e., the extent of deliberation as opposed to intuition (Capraro, 2019). While for the interaction structure there is a substantial consensus that sparse and heavily clustered networks help the spread of cooperation (Nowak, 2006; Ohtsuki et al., 2006), for the mode of cognition results are more articulated and depend on specific features of the social dilemma (Bear et al., 2017; Bear and Rand, 2016) and of the cost of deliberation (Jagau and van Veelen, 2017). An important aspect in evolutionary models is the behavioral rule adopted by agents, which heavily contributes to determining the trajectories of the dynamic adjustment. While the literature has extensively considered behavioral rules encompassing best reply (Bilancini and Boncinelli, 2009) and imitation (Levine and Pesendorfer, 2007) as well as processes of the type death-birth or birth-death (Ohtsuki et al., 2006), little attention has been given to evolutionary dynamics based on reinforcement learning (Tanabe and Masuda, 2012). Reinforcement learning is a prominent behavioral rule originated in behavioral sciences (Skinner, 1938a,b) and recently become extremely popular in computer sciences, with many different applications (Nian et al., 2020).

In recent years, increasing attention has been directed to leveraging pre-trained vision models for motor control. While existing works mainly emphasize the importance of this pre-training phase, the arguably equally important role played by downstream policy learning during control-specific fine-tuning is often neglected. It thus remains unclear if pre-trained vision models are consistent in their effectiveness under different control policies. To bridge this gap in understanding, we conduct a comprehensive study on 14 pre-trained vision models using 3 distinct classes of policy learning methods, including reinforcement learning (RL), imitation learning through behavior cloning (BC), and imitation learning with a visual reward function (VRF). Our study yields a series of intriguing results, including the discovery that the effectiveness of pre-training is highly dependent on the choice of the downstream policy learning algorithm. We show that conventionally accepted evaluation based on RL methods is highly variable and therefore unreliable, and further advocate for using more robust methods like VRF and BC. To facilitate more universal evaluations of pre-trained models and their policy learning methods in the future, we also release a benchmark of 21 tasks across 3 different environments alongside our work.

We revisit the standard formulation of tabular actor-critic algorithm as a two time-scale stochastic approximation with value function computed on a faster time-scale and policy computed on a slower time-scale. We observe that reversal of the time scales will in fact emulate value iteration and is a legitimate algorithm. We provide a proof of convergence and compare the two empirically with and without function approximation (with both linear and nonlinear function approximators) and observe that our proposed critic-actor algorithm performs on par with actor-critic in terms of both accuracy and computational effort. The actor-critic algorithm of Barto et al. [1] is one of the foremost reinforcement learning algorithms for data-driven approximate dynamic programming for Markov decision processes. Its rigorous analysis as a two time-scale stochastic approximation was initiated in [14] and [15], first in tabular form, then with linear function approximation, respectively.

Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We demonstrate that existing observation-space attacks on reinforcement learning agents have a common weakness: while effective, their lack of temporal consistency makes them detectable using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. We introduce perfect illusory attacks, a novel form of adversarial attack on sequential decision-makers that is both effective and provably statistically undetectable. We then propose the more versatile R-attacks, which result in observation transitions that are consistent with the state-transition function of the adversary-free environment and can be learned end-to-end. Compared to existing attacks, we empirically find R-attacks to be significantly harder to detect with automated methods, and a small study with human subjects suggests they are similarly harder to detect for humans. We propose that undetectability should be a central concern in the study of adversarial attacks on mixed-autonomy settings.

The dangers of adversarial attacks on Uncrewed Aerial Vehicle (UAV) agents operating in public are increasing. Adopting AI-based techniques and, more specifically, Deep Learning (DL) approaches to control and guide these UAVs can be beneficial in terms of performance but can add concerns regarding the safety of those techniques and their vulnerability against adversarial attacks. Confusion in the agent's decision-making process caused by these attacks can seriously affect the safety of the UAV. This paper proposes an innovative approach based on the explainability of DL methods to build an efficient detector that will protect these DL schemes and the UAVs adopting them from attacks. The agent adopts a Deep Reinforcement Learning (DRL) scheme for guidance and planning. The agent is trained with a Deep Deterministic Policy Gradient (DDPG) with Prioritised Experience Replay (PER) DRL scheme that utilises Artificial Potential Field (APF) to improve training times and obstacle avoidance performance. A simulated environment for UAV explainable DRL-based planning and guidance, including obstacles and adversarial attacks, is built. The adversarial attacks are generated by the Basic Iterative Method (BIM) algorithm and reduced obstacle course completion rates from 97\% to 35\%. Two adversarial attack detectors are proposed to counter this reduction. The first one is a Convolutional Neural Network Adversarial Detector (CNN-AD), which achieves accuracy in the detection of 80\%. The second detector utilises a Long Short Term Memory (LSTM) network. It achieves an accuracy of 91\% with faster computing times compared to the CNN-AD, allowing for real-time adversarial detection.

In multi-agent reinforcement learning, the use of a global objective is a powerful tool for incentivising cooperation. Unfortunately, it is not sample-efficient to train individual agents with a global reward, because it does not necessarily correlate with an agent's individual actions. This problem can be solved by factorising the global value function into local value functions. Early work in this domain performed factorisation by conditioning local value functions purely on local information. Recently, it has been shown that providing both local information and an encoding of the global state can promote cooperative behaviour. In this paper we propose QGNN, the first value factorisation method to use a graph neural network (GNN) based model. The multi-layer message passing architecture of QGNN provides more representational complexity than models in prior work, allowing it to produce a more effective factorisation. QGNN also introduces a permutation invariant mixer which is able to match the performance of other methods, even with significantly fewer parameters. We evaluate our method against several baselines, including QMIX-Att, GraphMIX, QMIX, VDN, and hybrid architectures. Our experiments include Starcraft, the standard benchmark for credit assignment; Estimate Game, a custom environment that explicitly models inter-agent dependencies; and Coalition Structure Generation, a foundational problem with real-world applications. The results show that QGNN outperforms state-of-the-art value factorisation baselines consistently.

Optimal control of general nonlinear systems is a central challenge in automation. Enabled by powerful function approximators, data-driven approaches to control have recently successfully tackled challenging applications. However, such methods often obscure the structure of dynamics and control behind black-box over-parameterized representations, thus limiting our ability to understand closed-loop behavior. This paper adopts a hybrid-system view of nonlinear modeling and control that lends an explicit hierarchical structure to the problem and breaks down complex dynamics into simpler localized units. We consider a sequence modeling paradigm that captures the temporal structure of the data and derive an expectation-maximization (EM) algorithm that automatically decomposes nonlinear dynamics into stochastic piecewise affine models with nonlinear transition boundaries. Furthermore, we show that these time-series models naturally admit a closed-loop extension that we use to extract local polynomial feedback controllers from nonlinear experts via behavioral cloning. Finally, we introduce a novel hybrid relative entropy policy search (Hb-REPS) technique that incorporates the hierarchical nature of hybrid models and optimizes a set of time-invariant piecewise feedback controllers derived from a piecewise polynomial approximation of a global state-value function.