In the autonomous driving field, the fusion of human knowledge into Deep Reinforcement Learning (DRL) is often based on the human demonstration recorded in the simulated environment. This limits the generalization and the feasibility of application in real-world traffic. We proposed a two-stage DRL method, that learns from real-world human driving to achieve performance that is superior to the pure DRL agent. Training a DRL agent is done within a framework for CARLA with Robot Operating System (ROS). For evaluation, we designed different real-world driving scenarios to compare the proposed two-stage DRL agent with the pure DRL agent. After extracting the 'good' behavior from the human driver, such as anticipation in a signalized intersection, the agent becomes more efficient and drives safer, which makes this autonomous agent more adapt to Human-Robot Interaction (HRI) traffic.

We study synchronous Q-learning with Polyak-Ruppert averaging (a.k.a., averaged Q-leaning) in a $\gamma$-discounted MDP. We establish asymptotic normality for the averaged iteration $\bar{\boldsymbol{Q}}_T$. Furthermore, we show that $\bar{\boldsymbol{Q}}_T$ is actually a regular asymptotically linear (RAL) estimator for the optimal Q-value function $\boldsymbol{Q}^*$ with the most efficient influence function. It implies the averaged Q-learning iteration has the smallest asymptotic variance among all RAL estimators. In addition, we present a non-asymptotic analysis for the $\ell_{\infty}$ error $\mathbb{E}\|\bar{\boldsymbol{Q}}_T-\boldsymbol{Q}^*\|_{\infty}$, showing it matches the instance-dependent lower bound as well as the optimal minimax complexity lower bound. As a byproduct, we find the Bellman noise has sub-Gaussian coordinates with variance $\mathcal{O}((1-\gamma)^{-1})$ instead of the prevailing $\mathcal{O}((1-\gamma)^{-2})$ under the standard bounded reward assumption. The sub-Gaussian result has potential to improve the sample complexity of many RL algorithms. In short, our theoretical analysis shows averaged Q-Leaning is statistically efficient.

State of the art deep reinforcement learning algorithms are sample inefficient due to the large number of episodes they require to achieve asymptotic performance. Episodic Reinforcement Learning (ERL) algorithms, inspired by the mammalian hippocampus, typically use extended memory systems to bootstrap learning from past events to overcome this sample-inefficiency problem. However, such memory augmentations are often used as mere buffers, from which isolated past experiences are drawn to learn from in an offline fashion (e.g., replay). Here, we demonstrate that including a bias in the acquired memory content derived from the order of episodic sampling improves both the sample and memory efficiency of an episodic control algorithm. We test our Sequential Episodic Control (SEC) model in a foraging task to show that storing and using integrated episodes as event sequences leads to faster learning with fewer memory requirements as opposed to a standard ERL benchmark, Model-Free Episodic Control, that buffers isolated events only. We also study the effect of memory constraints and forgetting on the sequential and non-sequential version of the SEC algorithm. Furthermore, we discuss how a hippocampal-like fast memory system could bootstrap slow cortical and subcortical learning subserving habit formation in the mammalian brain.

This work investigates the effects of Curriculum Learning (CL)-based approaches on the agent's performance. In particular, we focus on the safety aspect of robotic mapless navigation, comparing over a standard end-to-end (E2E) training strategy. To this end, we present a CL approach that leverages Transfer of Learning (ToL) and fine-tuning in a Unity-based simulation with the Robotnik Kairos as a robotic agent. For a fair comparison, our evaluation considers an equal computational demand for every learning approach (i.e., the same number of interactions and difficulty of the environments) and confirms that our CL-based method that uses ToL outperforms the E2E methodology. In particular, we improve the average success rate and the safety of the trained policy, resulting in 10% fewer collisions in unseen testing scenarios. To further confirm these results, we employ a formal verification tool to quantify the number of correct behaviors of Reinforcement Learning policies over desired specifications.

This course is about Reinforcement Learning. The first step is to talk about the mathematic. This course is about Reinforcement Learning. The first step is to talk about the mathematical background: we can use a Markov Decision Process as a model for reinforcement learning. We can solve the problem 3 ways: value-iteration, policy-iteration and Q-learning.

In dynamic programming (DP) and reinforcement learning (RL), an agent learns to act optimally in terms of expected long-term return by sequentially interacting with its environment modeled by a Markov decision process (MDP). More generally in distributional reinforcement learning (DRL), the focus is on the whole distribution of the return, not just its expectation. Although DRL-based methods produced state-of-the-art performance in RL with function approximation, they involve additional quantities (compared to the non-distributional setting) that are still not well understood. As a first contribution, we introduce a new class of distributional operators, together with a practical DP algorithm for policy evaluation, that come with a robust MDP interpretation. Indeed, our approach reformulates through an augmented state space where each state is split into a worst-case substate and a best-case substate, whose values are maximized by safe and risky policies respectively. Finally, we derive distributional operators and DP algorithms solving a new control task: How to distinguish safe from risky optimal actions in order to break ties in the space of optimal policies?

Recent works have shown that the computational efficiency of video recognition can be significantly improved by reducing the spatial redundancy. As a representative work, the adaptive focus method (AdaFocus) has achieved a favorable trade-off between accuracy and inference speed by dynamically identifying and attending to the informative regions in each video frame. However, AdaFocus requires a complicated three-stage training pipeline (involving reinforcement learning), leading to slow convergence and is unfriendly to practitioners. This work reformulates the training of AdaFocus as a simple one-stage algorithm by introducing a differentiable interpolation-based patch selection operation, enabling efficient end-to-end optimization. We further present an improved training scheme to address the issues introduced by the one-stage formulation, including the lack of supervision, input diversity and training stability. Moreover, a conditional-exit technique is proposed to perform temporal adaptive computation on top of AdaFocus without additional training. Extensive experiments on six benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, and Jester) demonstrate that our model significantly outperforms the original AdaFocus and other competitive baselines, while being considerably more simple and efficient to train. Code is available at https://github.com/LeapLabTHU/AdaFocusV2.

Mobile edge computing has become an effective and fundamental paradigm for futuristic autonomous vehicles to offload computing tasks. However, due to the high mobility of vehicles, the dynamics of the wireless conditions, and the uncertainty of the arrival computing tasks, it is difficult for a single vehicle to determine the optimal offloading strategy. In this paper, we propose a Digital Twin (DT) empowered task offloading framework for Internet of Vehicles. As a software agent residing in the cloud, a DT can obtain both global network information by using communications among DTs, and historical information of a vehicle by using the communications within the twin. The global network information and historical vehicular information can significantly facilitate the offloading. In specific, to preserve the precious computing resource at different levels for most appropriate computing tasks, we integrate a learning scheme based on the prediction of futuristic computing tasks in DT. Accordingly, we model the offloading scheduling process as a Markov Decision Process (MDP) to minimize the long-term cost in terms of a trade off between task latency, energy consumption, and renting cost of clouds. Simulation results demonstrate that our algorithm can effectively find the optimal offloading strategy, as well as achieve the fast convergence speed and high performance, compared with other existing approaches.

This paper studies the regret minimization problem for finite horizon, episodic reinforcement learning (RL) with infinitely large state and action spaces. Empirically, RL has achieved success in diverse domains, even when the problem size (measured in the number of states and actions) explodes [35, 44, 28]. The key to developing sample-efficient algorithms is to leverage function approximation, enabling us to generalize across different state-action pairs. Much theoretical progress has been made towards understanding function approximation in RL. Existing theory typically requires strong linearity assumptions on transition dynamics [e.g., 55, 26, 10, 36] or action-value functions [e.g., 30, 57] of the Markov Decision Process (MDP). However, most real world problems are nonlinear, and our theoretical understanding of these settings remains limited. Thus, we ask the question: Can we design provably efficient RL algorithms in nonlinear environments? Recently, Chowdhury et al. [13] introduced a nonlinear setting where the state-transition measures are finitely parameterized exponential family models, and they proposed to estimate model parameters via maximum likelihood estimation (MLE). The exponential family is a well-studied and powerful statistical framework, so it is a natural model class to consider beyond linear models.

We consider large-scale Markov decision processes with an unknown cost function and address the problem of learning a policy from a finite set of expert demonstrations. We assume that the learner is not allowed to interact with the expert and has no access to reinforcement signal of any kind. Existing inverse reinforcement learning methods come with strong theoretical guarantees, but are computationally expensive, while state-of-the-art policy optimization algorithms achieve significant empirical success, but are hampered by limited theoretical understanding. To bridge the gap between theory and practice, we introduce a novel bilinear saddle-point framework using Lagrangian duality. The proposed primal-dual viewpoint allows us to develop a model-free provably efficient algorithm through the lens of stochastic convex optimization. The method enjoys the advantages of simplicity of implementation, low memory requirements, and computational and sample complexities independent of the number of states. We further present an equivalent no-regret online-learning interpretation.