A dialogue policy module is an essential part of task-completion dialogue systems. Recently, increasing interest has focused on reinforcement learning (RL)-based dialogue policy. Its favorable performance and wise action decisions rely on an accurate estimation of action values. The overestimation problem is a widely known issue of RL since its estimate of the maximum action value is larger than the ground truth, which results in an unstable learning process and suboptimal policy. This problem is detrimental to RL-based dialogue policy learning. To mitigate this problem, this paper proposes a dynamic partial average estimator (DPAV) of the ground truth maximum action value. DPAV calculates the partial average between the predicted maximum action value and minimum action value, where the weights are dynamically adaptive and problem-dependent. We incorporate DPAV into a deep Q-network as the dialogue policy and show that our method can achieve better or comparable results compared to top baselines on three dialogue datasets of different domains with a lower computational load. In addition, we also theoretically prove the convergence and derive the upper and lower bounds of the bias compared with those of other methods.

Deepmind has been the pioneer in making AI models that have the capability to mimic a human's cognitive ability to play games. Games are a common testbed to assess a model's ability. After mastering games like Go, Chess and Checkers, Deepmind has launched DeepNash, an AI model that can play Stratego at an expert level. Mastering a game like'Stratego' is a significant achievement for AI research because it presents a challenging benchmark for learning strategic interactions at a massive scale. Stratego's complexity is based on two key aspects. Firstly, there are 10535 possible states in the game, which is exponentially larger than Texas hold'em poker(10164 states) and Go(10360 states).

We study policy gradient (PG) for reinforcement learning in continuous time and space under the regularized exploratory formulation developed by Wang et al. (2020). We represent the gradient of the value function with respect to a given parameterized stochastic policy as the expected integration of an auxiliary running reward function that can be evaluated using samples and the current value function. This representation effectively turns PG into a policy evaluation (PE) problem, enabling us to apply the martingale approach recently developed by Jia and Zhou (2022a) for PE to solve our PG problem. Based on this analysis, we propose two types of actor-critic algorithms for RL, where we learn and update value functions and policies simultaneously and alternatingly. The first type is based directly on the aforementioned representation, which involves future trajectories and is offline. The second type, designed for online learning, employs the first-order condition of the policy gradient and turns it into martingale orthogonality conditions. These conditions are then incorporated using stochastic approximation when updating policies. Finally, we demonstrate the algorithms by simulations in two concrete examples.

Reinforcement learning (RL) has shown to reach super human-level performance across a wide range of tasks. However, unlike supervised machine learning, learning strategies that generalize well to a wide range of situations remains one of the most challenging problems for real-world RL. Autonomous driving (AD) provides a multi-faceted experimental field, as it is necessary to learn the correct behavior over many variations of road layouts and large distributions of possible traffic situations, including individual driver personalities and hard-to-predict traffic events. In this paper we propose a challenging benchmark for generalizable RL for AD based on a configurable, flexible, and performant code base. Our benchmark uses a catalog of randomized scenario generators, including multiple mechanisms for road layout and traffic variations, different numerical and visual observation types, distinct action spaces, diverse vehicle models, and allows for use under static scenario definitions. In addition to purely algorithmic insights, our application-oriented benchmark also enables a better understanding of the impact of design decisions such as action and observation space on the generalizability of policies. Our benchmark aims to encourage researchers to propose solutions that are able to successfully generalize across scenarios, a task in which current RL methods fail.

As an emerging technology, Connected Autonomous Vehicles (CAVs) are believed to have the ability to move through intersections in a faster and safer manner, through effective Vehicle-to-Everything (V2X) communication and global observation. Autonomous intersection management is a key path to efficient crossing at intersections, which reduces unnecessary slowdowns and stops through adaptive decision process of each CAV, enabling fuller utilization of the intersection space. Distributed reinforcement learning (DRL) offers a flexible, end-to-end model for AIM, adapting for many intersection scenarios. While DRL is prone to collisions as the actions of multiple sides in the complicated interactions are sampled from a generic policy, restricting the application of DRL in realistic scenario. To address this, we propose a hierarchical RL framework where models at different levels vary in receptive scope, action step length, and feedback period of reward. The upper layer model accelerate CAVs to prevent them from being clashed, while the lower layer model adjust the trends from upper layer model to avoid the change of mobile state causing new conflicts. And the real action of CAV at each step is co-determined by the trends from both levels, forming a real-time balance in the adversarial process. The proposed model is proven effective in the experiment undertaken in a complicated intersection with 4 branches and 4 lanes each branch, and show better performance compared with baselines.

A two-wheeled self-balancing robot is an example of an inverse pendulum and is an inherently non-linear, unstable system. The fundamental concept of the proposed framework "Epersist" is to overcome the challenge of counterbalancing an initially unstable system by delivering robust control mechanisms, Proportional Integral Derivative(PID), and Reinforcement Learning (RL). Moreover, the micro-controller NodeMCUESP32 and inertial sensor in the Epersist employ fewer computational procedures to give accurate instruction regarding the spin of wheels to the motor driver, which helps control the wheels and balance the robot. This framework also consists of the mathematical model of the PID controller and a novel self-trained advantage actor-critic algorithm as the RL agent. After several experiments, control variable calibrations are made as the benchmark values to attain the angle of static equilibrium. This "Epersist" framework proposes PID and RL-assisted functional prototypes and simulations for better utility.

Developing robust vision-guided controllers for quadrupedal robots in complex environments, with various obstacles, dynamical surroundings and uneven terrains, is very challenging. While Reinforcement Learning (RL) provides a promising paradigm for agile locomotion skills with vision inputs in simulation, it is still very challenging to deploy the RL policy in the real world. Our key insight is that aside from the discrepancy in the domain gap, in visual appearance between the simulation and the real world, the latency from the control pipeline is also a major cause of difficulty. In this paper, we propose Multi-Modal Delay Randomization (MMDR) to address this issue when training RL agents. Specifically, we simulate the latency of real hardware by using past observations, sampled with randomized periods, for both proprioception and vision. We train the RL policy for end-to-end control in a physical simulator without any predefined controller or reference motion, and directly deploy it on the real A1 quadruped robot running in the wild. We evaluate our method in different outdoor environments with complex terrains and obstacles. We demonstrate the robot can smoothly maneuver at a high speed, avoid the obstacles, and show significant improvement over the baselines. Our project page with videos is at https://mehooz.github.io/mmdr-wild/.

The placement of charging stations in areas with developing charging infrastructure is a critical component of the future success of electric vehicles (EVs). In Albany County in New York, the expected rise in the EV population requires additional charging stations to maintain a sufficient level of efficiency across the charging infrastructure. A novel application of Reinforcement Learning (RL) is able to find optimal locations for new charging stations given the predicted charging demand and current charging locations. The most important factors that influence charging demand prediction include the conterminous traffic density, EV registrations, and proximity to certain types of public buildings. The proposed RL framework can be refined and applied to cities across the world to optimize charging station placement.

The combination of learning methods with Model Predictive Control (MPC) has attracted a significant amount of attention in the recent literature. The hope of this combination is to reduce the reliance of MPC schemes on accurate models, and to tap into the fast developing machine learning and reinforcement learning tools to exploit the growing amount of data available for many systems. In particular, the combination of reinforcement learning and MPC has been proposed as a viable and theoretically justified approach to introduce explainable, safe and stable policies in reinforcement learning. However, a formal theory detailing how the safety and stability of an MPC-based policy can be maintained through the parameter updates delivered by the learning tools is still lacking. This paper addresses this gap. The theory is developed for the generic Robust MPC case, and applied in simulation in the robust tube-based linear MPC case, where the theory is fairly easy to deploy in practice. The paper focuses on Reinforcement Learning as a learning tool, but it applies to any learning method that updates the MPC parameters online.

This paper studies policy optimization algorithms for multi-agent reinforcement learning. We begin by proposing an algorithm framework for two-player zero-sum Markov Games in the full-information setting, where each iteration consists of a policy update step at each state using a certain matrix game algorithm, and a value update step with a certain learning rate. This framework unifies many existing and new policy optimization algorithms. We show that the state-wise average policy of this algorithm converges to an approximate Nash equilibrium (NE) of the game, as long as the matrix game algorithms achieve low weighted regret at each state, with respect to weights determined by the speed of the value updates. Next, we show that this framework instantiated with the Optimistic Follow-The-Regularized-Leader (OFTRL) algorithm at each state (and smooth value updates) can find an $\mathcal{\widetilde{O}}(T^{-5/6})$ approximate NE in $T$ iterations, and a similar algorithm with slightly modified value update rule achieves a faster $\mathcal{\widetilde{O}}(T^{-1})$ convergence rate. These improve over the current best $\mathcal{\widetilde{O}}(T^{-1/2})$ rate of symmetric policy optimization type algorithms. We also extend this algorithm to multi-player general-sum Markov Games and show an $\mathcal{\widetilde{O}}(T^{-3/4})$ convergence rate to Coarse Correlated Equilibria (CCE). Finally, we provide a numerical example to verify our theory and investigate the importance of smooth value updates, and find that using "eager" value updates instead (equivalent to the independent natural policy gradient algorithm) may significantly slow down the convergence, even on a simple game with $H=2$ layers.