Deep reinforcement learning (DRL) integrates the feature representation ability of deep learning with the decision-making ability of reinforcement learning so that it can achieve powerful end-to-end learning control capabilities. In the past decade, DRL has made substantial advances in many tasks that require perceiving high-dimensional input and making optimal or near-optimal decisions. However, there are still many challenging problems in the theory and applications of DRL, especially in learning control tasks with limited samples, sparse rewards, and multiple agents. Researchers have proposed various solutions and new theories to solve these problems and promote the development of DRL. In addition, deep learning has stimulated the further development of many subfields of reinforcement learning, such as hierarchical reinforcement learning (HRL), multiagent reinforcement learning, and imitation learning. This article gives a comprehensive overview of the fundamental theories, key algorithms, and primary research domains of DRL. In addition to value-based and policy-based DRL algorithms, the advances in maximum entropy-based DRL are summarized. The future research topics of DRL are also analyzed and discussed.

In this paper, we present a kernel-based least squares policy iteration (KLSPI) algorithm for reinforcement learning (RL) in large or continuous state spaces, which can be used to realize adaptive feedback control of uncertain dynamic systems. By using KLSPI, near-optimal control policies can be obtained without much a priori knowledge on dynamic models of control plants. In KLSPI, Mercer kernels are used in the policy evaluation of a policy iteration process, where a new kernel-based least squares temporal-difference algorithm called KLSTD-Q is proposed for efficient policy evaluation. To keep the sparsity and improve the generalization ability of KLSTD-Q solutions, a kernel sparsification procedure based on approximate linear dependency (ALD) is performed. Compared to the previous works on approximate RL methods, KLSPI makes two progresses to eliminate the main difficulties of existing results. One is the better convergence and (near) optimality guarantee by using the KLSTD-Q algorithm for policy evaluation with high precision. The other is the automatic feature selection using the ALD-based kernel sparsification. Therefore, the KLSPI algorithm provides a general RL method with generalization performance and convergence guarantee for large-scale Markov decision problems (MDPs). Experimental results on a typical RL task for a stochastic chain problem demonstrate that KLSPI can consistently achieve better learning efficiency and policy quality than the previous least squares policy iteration (LSPI) algorithm. Furthermore, the KLSPI method was also evaluated on two nonlinear feedback control problems, including a ship heading control problem and the swing up control of a double-link underactuated pendulum called acrobot. Simulation results illustrate that the proposed method can optimize controller performance using little a priori information of uncertain dynamic systems. It is also demonstrated that KLSPI can be applied to online learning control by incorporating an initial controller to ensure online performance.

Autonomous driving technologies have received notable attention in the past decades. In autonomous driving systems, identifying a precise dynamical model for motion control is nontrivial due to the strong nonlinearity and uncertainty in vehicle dynamics. Recent efforts have resorted to machine learning techniques for building vehicle dynamical models, but the generalization ability and interpretability of existing methods still need to be improved. In this paper, we propose a pure data-driven vehicle modeling approach based on deep neural networks with an interpretable Koopman operator. The main advantage of using the Koopman operator is to represent the nonlinear dynamics in a linear lifted feature space. In the proposed approach, a deep learning-based extended dynamic mode decomposition algorithm is presented to learn a finite-dimensional approximation of the Koopman operator. A multi-step prediction loss function is used in the training process, enabling a long-term prediction capability. Furthermore, a data-driven model predictive controller with the learned Koopman model is designed for velocity profile tracking control of autonomous vehicles. Simulation results in a high-fidelity CarSim environment show that our approach outperforms previously developed traditional and advanced modeling methods. Velocity profile tracking tests of the autonomous vehicle are also performed in the CarSim environment. The results show that our approach has better tracking accuracy and higher computational efficiency than the model predictive control algorithms using a nonlinear model and a linear time-varying model.

Autonomous decision making is a critical and difficult task for intelligent vehicles in dynamic transportation environments. In this paper, a reinforcement learning approach with value function approximation and feature learning is proposed for autonomous decision making of intelligent vehicles on highways. In the proposed approach, the sequential decision making problem for lane changing and overtaking is modeled as a Markov decision process with multiple goals, including safety, speediness, smoothness, etc. In order to learn optimized policies for autonomous decision-making, a multiobjective approximate policy iteration (MO-API) algorithm is presented. The features for value function approximation are learned in a data-driven way, where sparse kernel-based features or manifold-based features can be constructed based on data samples. Compared with previous RL algorithms such as multiobjective Q-learning, the MO-API approach uses data-driven feature representation for value and policy approximation so that better learning efficiency can be achieved. A highway simulation environment using a 14 degree-of-freedom vehicle dynamics model was established to generate training data and test the performance of different decision-making methods for intelligent vehicles on highways. The results illustrate the advantages of the proposed MO-API method under different traffic conditions. Furthermore, we also tested the learned decision policy on a real autonomous vehicle to implement overtaking decision and control under normal traffic on highways. The experimental results also demonstrate the effectiveness of the proposed method.

Multiple unmanned aerial vehicles (UAVs) are able to efficiently accomplish a variety of tasks in complex scenarios. However, developing a collision-avoiding flocking policy for multiple fixed-wing UAVs is still challenging, especially in obstacle-cluttered environments. In this article, we propose a novel curriculum-based multiagent deep reinforcement learning (MADRL) approach called task-specific curriculum-based MADRL (TSCAL) to learn the decentralized flocking with obstacle avoidance policy for multiple fixed-wing UAVs. The core idea is to decompose the collision-avoiding flocking task into multiple subtasks and progressively increase the number of subtasks to be solved in a staged manner. Meanwhile, TSCAL iteratively alternates between the procedures of online learning and offline transfer. For online learning, we propose a hierarchical recurrent attention multiagent actor–critic (HRAMA) algorithm to learn the policies for the corresponding subtask(s) in each learning stage. For offline transfer, we develop two transfer mechanisms, i.e., model reload and buffer reuse, to transfer knowledge between two neighboring stages. A series of numerical simulations demonstrate the significant advantages of TSCAL in terms of policy optimality, sample efficiency, and learning stability. Finally, the high-fidelity hardware-in-the-loop (HITL) simulation is conducted to verify the adaptability of TSCAL. A video about the numerical and HITL simulations is available at https://youtu.be/R9yLJNYRIqY.