The study of decentralized learning or independent learning in cooperative multi-agent reinforcement learning has a history of decades. Recently empirical studies show that independent PPO (IPPO) can obtain good performance, close to or even better than the methods of centralized training with decentralized execution, in several benchmarks. However, decentralized actor-critic with convergence guarantee is still open. In this paper, we propose \textit{decentralized policy optimization} (DPO), a decentralized actor-critic algorithm with monotonic improvement and convergence guarantee. We derive a novel decentralized surrogate for policy optimization such that the monotonic improvement of joint policy can be guaranteed by each agent \textit{independently} optimizing the surrogate. In practice, this decentralized surrogate can be realized by two adaptive coefficients for policy optimization at each agent. Empirically, we compare DPO with IPPO in a variety of cooperative multi-agent tasks, covering discrete and continuous action spaces, and fully and partially observable environments. The results show DPO outperforms IPPO in most tasks, which can be the evidence for our theoretical results.

In this short note we derive a relationship between the Bregman divergence from the current policy to the optimal policy and the suboptimality of the current value function in a regularized Markov decision process. This result has implications for multi-task reinforcement learning, offline reinforcement learning, and regret analysis under function approximation, among others. The main result of this manuscript holds more generally, but for brevity we shall restrict ourselves to this case. To prove our main result we require a slight generalization of the performance difference lemma (PDL)[1] to cover the regularized MDP case. The proof of this identity is included in the appendix for completeness.

Proper functioning of connected and automated vehicles (CAVs) is crucial for the safety and efficiency of future intelligent transport systems. Meanwhile, transitioning to fully autonomous driving requires a long period of mixed autonomy traffic, including both CAVs and human-driven vehicles. Thus, collaboration decision-making for CAVs is essential to generate appropriate driving behaviors to enhance the safety and efficiency of mixed autonomy traffic. In recent years, deep reinforcement learning (DRL) has been widely used in solving decision-making problems. However, the existing DRL-based methods have been mainly focused on solving the decision-making of a single CAV. Using the existing DRL-based methods in mixed autonomy traffic cannot accurately represent the mutual effects of vehicles and model dynamic traffic environments. To address these shortcomings, this article proposes a graph reinforcement learning (GRL) approach for multi-agent decision-making of CAVs in mixed autonomy traffic. First, a generic and modular GRL framework is designed. Then, a systematic review of DRL and GRL methods is presented, focusing on the problems addressed in recent research. Moreover, a comparative study on different GRL methods is further proposed based on the designed framework to verify the effectiveness of GRL methods. Results show that the GRL methods can well optimize the performance of multi-agent decision-making for CAVs in mixed autonomy traffic compared to the DRL methods. Finally, challenges and future research directions are summarized. This study can provide a valuable research reference for solving the multi-agent decision-making problems of CAVs in mixed autonomy traffic and can promote the implementation of GRL-based methods into intelligent transportation systems. The source code of our work can be found at https://github.com/Jacklinkk/Graph_CAVs.

State abstraction and function approximation are vital components used by reinforcement learning (RL) algorithms to efficiently solve complex control problems when exact computations are intractable due to large state and action spaces. Over the past few decades, state abstraction in RL has evolved from the use of pre-determined and problemspecific features [18, 74, 9, 69, 64, 42, 58] to the use of adaptive basis functions learnt by solving an isolated regression problem [53, 47, 39, 56], and more recently to the use of neural network-based Deep-RL algorithms that embed state abstraction in successive layers of a neural network [5, 7]. Feature abstraction results in information loss, and the resulting state features might not satisfy the controlled Markov property, even if this property is satisfied by the corresponding state [70]. One approach to counteract the loss of the Markov property is to generate the features using the history of state-action pairs, and empirical evidence suggests that using such history-based features are beneficial in practice [52]. However, a theoretical characterisation of history-based Deep-RL algorithms for fully observed Markov Decision Processes (MDPs) is largely absent form the literature.

We investigate the problem of risk averse robot path planning using the deep reinforcement learning and distributionally robust optimization perspectives. Our problem formulation involves modelling the robot as a stochastic linear dynamical system, assuming that a collection of process noise samples is available. We cast the risk averse motion planning problem as a Markov decision process and propose a continuous reward function design that explicitly takes into account the risk of collision with obstacles while encouraging the robot's motion towards the goal. We learn the risk-averse robot control actions through Lipschitz approximated Wasserstein distributionally robust deep Q-learning to hedge against the noise uncertainty. The learned control actions result in a safe and risk averse trajectory from the source to the goal, avoiding all the obstacles. Various supporting numerical simulations are presented to demonstrate our proposed approach.

The explosive growth of dynamic and heterogeneous data traffic brings great challenges for 5G and beyond mobile networks. To enhance the network capacity and reliability, we propose a learning-based dynamic time-frequency division duplexing (D-TFDD) scheme that adaptively allocates the uplink and downlink time-frequency resources of base stations (BSs) to meet the asymmetric and heterogeneous traffic demands while alleviating the inter-cell interference. We formulate the problem as a decentralized partially observable Markov decision process (Dec-POMDP) that maximizes the long-term expected sum rate under the users' packet dropping ratio constraints. In order to jointly optimize the global resources in a decentralized manner, we propose a federated reinforcement learning (RL) algorithm named federated Wolpertinger deep deterministic policy gradient (FWDDPG) algorithm. The BSs decide their local time-frequency configurations through RL algorithms and achieve global training via exchanging local RL models with their neighbors under a decentralized federated learning framework. Specifically, to deal with the large-scale discrete action space of each BS, we adopt a DDPG-based algorithm to generate actions in a continuous space, and then utilize Wolpertinger policy to reduce the mapping errors from continuous action space back to discrete action space. Simulation results demonstrate the superiority of our proposed algorithm to benchmark algorithms with respect to system sum rate.

Cooperative multi-agent reinforcement learning (MARL) has made prominent progress in recent years. For training efficiency and scalability, most of the MARL algorithms make all agents share the same policy or value network. However, in many complex multi-agent tasks, different agents are expected to possess specific abilities to handle different subtasks. In those scenarios, sharing parameters indiscriminately may lead to similar behavior across all agents, which will limit the exploration efficiency and degrade the final performance. To balance the training complexity and the diversity of agent behavior, we propose a novel framework to learn dynamic subtask assignment (LDSA) in cooperative MARL. Specifically, we first introduce a subtask encoder to construct a vector representation for each subtask according to its identity. To reasonably assign agents to different subtasks, we propose an ability-based subtask selection strategy, which can dynamically group agents with similar abilities into the same subtask. In this way, agents dealing with the same subtask share their learning of specific abilities and different subtasks correspond to different specific abilities. We further introduce two regularizers to increase the representation difference between subtasks and stabilize the training by discouraging agents from frequently changing subtasks, respectively. Empirical results show that LDSA learns reasonable and effective subtask assignment for better collaboration and significantly improves the learning performance on the challenging StarCraft II micromanagement benchmark and Google Research Football.

Unlearn.AI, Inc., 75 Hawthorne St. Ste 560, San Francisco, CA 94105 (Dated: November 7, 2022) Restricted Boltzmann Machines (RBMs) are probabilistic generative models that can be trained by maximum likelihood in principle, but are usually trained by an approximate algorithm called Contrastive Divergence (CD) in practice. In general, a CD-k algorithm estimates an average with respect to the model distribution using a sample obtained from a k-step Markov Chain Monte Carlo Algorithm (e.g., block Gibbs sampling) starting from some initial configuration. Choices of k typically vary from 1 to 100. This technical report explores if it's possible to leverage a simple approximate sampling algorithm with a modified version of CD in order to train an RBM with k=0. As usual, the method is illustrated on MNIST.

We study the convergence of several natural policy gradient (NPG) methods in infinite-horizon discounted Markov decision processes with regular policy parametrizations. For a variety of NPGs and reward functions we show that the trajectories in state-action space are solutions of gradient flows with respect to Hessian geometries, based on which we obtain global convergence guarantees and convergence rates. In particular, we show linear convergence for unregularized and regularized NPG flows with the metrics proposed by Kakade and Morimura and co-authors by observing that these arise from the Hessian geometries of conditional entropy and entropy respectively. Further, we obtain sublinear convergence rates for Hessian geometries arising from other convex functions like log-barriers. Finally, we interpret the discrete-time NPG methods with regularized rewards as inexact Newton methods if the NPG is defined with respect to the Hessian geometry of the regularizer. This yields local quadratic convergence rates of these methods for step size equal to the penalization strength.

This paper is concerned with constructing a confidence interval for a target policy's value offline based on a pre-collected observational data in infinite horizon settings. Most of the existing works assume no unmeasured variables exist that confound the observed actions. This assumption, however, is likely to be violated in real applications such as healthcare and technological industries. In this paper, we show that with some auxiliary variables that mediate the effect of actions on the system dynamics, the target policy's value is identifiable in a confounded Markov decision process. Based on this result, we develop an efficient off-policy value estimator that is robust to potential model misspecification and provide rigorous uncertainty quantification. Our method is justified by theoretical results, simulated and real datasets obtained from ridesharing companies. A Python implementation of the proposed procedure is available at https://github.com/Mamba413/cope.