Over the years, Reinforcement Learning (RL) established itself as a convenient paradigm to learn optimal policies from data. However, most RL algorithms achieve optimal policies by exploring all the possible actions and this, in real-world scenarios, is often infeasible or impractical due to e.g. safety constraints. Motivated by this, in this paper we propose to augment RL with Model Predictive Control (MPC), a popular model-based control algorithm that allows to optimally control a system while satisfying a set of constraints. The result is an algorithm, the MPC-augmented RL algorithm (MPCaRL) that makes use of MPC to both drive how RL explores the actions and to modify the corresponding rewards. We demonstrate the effectiveness of the MPCaRL by letting it play against the Atari game Pong. The results obtained highlight the ability of the algorithm to learn general tasks with essentially no training.

The information bottleneck principle is an elegant and useful approach to representation learning. In this paper, we investigate the problem of representation learning in the context of reinforcement learning using the information bottleneck framework, aiming at improving the sample efficiency of the learning algorithms. %by accelerating the process of discarding irrelevant information when the %input states are extremely high-dimensional. We analytically derive the optimal conditional distribution of the representation, and provide a variational lower bound. Then, we maximize this lower bound with the Stein variational (SV) gradient method. We incorporate this framework in the advantageous actor critic algorithm (A2C) and the proximal policy optimization algorithm (PPO). Our experimental results show that our framework can improve the sample efficiency of vanilla A2C and PPO significantly. Finally, we study the information bottleneck (IB) perspective in deep RL with the algorithm called mutual information neural estimation(MINE) . We experimentally verify that the information extraction-compression process also exists in deep RL and our framework is capable of accelerating this process. We also analyze the relationship between MINE and our method, through this relationship, we theoretically derive an algorithm to optimize our IB framework without constructing the lower bound.

Quantum computing exploits basic quantum phenomena such as state superposition and entanglement to perform computations. The Quantum Approximate Optimization Algorithm (QAOA) is arguably one of the leading quantum algorithms that can outperform classical state-of-the-art methods in the near term. QAOA is a hybrid quantum-classical algorithm that combines a parameterized quantum state evolution with a classical optimization routine to approximately solve combinatorial problems. The quality of the solution obtained by QAOA within a fixed budget of calls to the quantum computer depends on the performance of the classical optimization routine used to optimize the variational parameters. In this work, we propose an approach based on reinforcement learning (RL) to train a policy network that can be used to quickly find high-quality variational parameters for unseen combinatorial problem instances. The RL agent is trained on small problem instances which can be simulated on a classical computer, yet the learned RL policy is generalizable and can be used to efficiently solve larger instances. Extensive simulations using the IBM Qiskit Aer quantum circuit simulator demonstrate that our trained RL policy can reduce the optimality gap by a factor up to 8.61 compared with other off-the-shelf optimizers tested.

We present the first provably convergent off-policy actor-critic algorithm (COF-PAC) with function approximation in a two-timescale form. Key to COF-PAC is the introduction of a new critic, the emphasis critic, which is trained via Gradient Emphasis Learning (GEM), a novel combination of the key ideas of Gradient Temporal Difference Learning and Emphatic Temporal Difference Learning. With the help of the emphasis critic and the canonical value function critic, we show convergence for COF-PAC, where the critics are linear and the actor can be nonlinear.

Answering complex questions involving multiple entities and relations is a challenging task. Logically, the answer to a complex question should be derived by decomposing the complex question into multiple simple sub-questions and then answering those sub-questions. Existing work has followed this strategy but has not attempted to optimize the order how those sub-questions are answered. As a result, the sub-questions are answered in an arbitrary order, leading to larger search space and higher risk of missing an answer. In this paper, we propose a novel reinforcement learning (RL) approach to answering complex questions that can learn a policy to dynamically decide which sub-question should be answered at each state of reasoning. We leverage the expected value-variance criterion to enable the learned policy to balance between the risk and utility of answering a sub-question. Experiment results show that the RL approach can substantially improve the optimal-ity of ordering the sub-questions, leading to improved accuracy of question answering. The proposed method for learning to order sub-questions is general and can thus be potentially combined with many existing ideas for answering complex questions to enhance their performance. Introduction Real-world questions can be complex, involving multiple interrelated entities and relations, which we refer to as complex questions . For example, "who writes Harry Potter" is a simple question that only involves a single entity and a relation, while "Which city is the filming location of the book written by J.K.Rowling and held Olympics?" is a complex question, which consists of multiple entities and relations. How to automatically answer such complex questions is a significant scientific challenge because it requires a system to capture the dependencies between different components of the questions and reason over them. Recently, some recent work has attempted to tackle such complex questions (Talmor and Berant 2018; Iyyer, Yih, and Chang 2016; Min et al. 2019; Zhang et al. 2019), usually by decomposing a complex question into a sequence of simple questions and answering them based on a computation tree derived from the original question that can capture the dependency between sub-questions as shown in Figure 1.

The capability to learn and adapt to changes in the driving environment is crucial for developing autonomous driving systems that are scalable beyond geo-fenced operational design domains. Deep Reinforcement Learning (RL) provides a promising and scalable framework for developing adaptive learning based solutions. Deep RL methods usually model the problem as a (Partially Observable) Markov Decision Process in which an agent acts in a stationary environment to learn an optimal behavior policy. However, driving involves complex interaction between multiple, intelligent (artificial or human) agents in a highly non-stationary environment. In this paper, we propose the use of Partially Observable Markov Games(POSG) for formulating the connected autonomous driving problems with realistic assumptions. We provide a taxonomy of multi-agent learning environments based on the nature of tasks, nature of agents and the nature of the environment to help in categorizing various autonomous driving problems that can be addressed under the proposed formulation. As our main contributions, we provide MACAD-Gym, a Multi-Agent Connected, Autonomous Driving agent learning platform for furthering research in this direction. Our MACAD-Gym platform provides an extensible set of Connected Autonomous Driving (CAD) simulation environments that enable the research and development of Deep RL- based integrated sensing, perception, planning and control algorithms for CAD systems with unlimited operational design domain under realistic, multi-agent settings. We also share the MACAD-Agents that were trained successfully using the MACAD-Gym platform to learn control policies for multiple vehicle agents in a partially observable, stop-sign controlled, 3-way urban intersection environment with raw (camera) sensor observations.

The basic actor-critic model stems from Deep Convolution Q-Learning which is where the agent implements q-learning, but instead of taking in a matrix of states as input, it takes in images and feeds them into a deep convolutional neural network. Don't worry about the rectangles on the right side, they represent a deep neural network with all the nodes and connections. It's just easier to explain and understand A3C this way. In a regular Deep Convolution Q-Learning network, there would only be one output and that would be the q-values of the different actions. However in A3C, there are two outputs, one of the q-values for the different actions and the other to calculate the value of being in the state the agent is actually in.

Game theory plays a fundamental factor in modern artificial intelligence(AI) solutions. Specifically, deep reinforcement learning(DRL) is an area of AI that embraced game theory as a first-class citize. From single-agent programs to complex multi-agent DRL environments, gamifying dynamics are present across the lifecycle of AI programs. The fascinating thing is that the rapid evolution of DRL has also triggered a renewed interesting in game theory research. The relationship between game theory and DRL seems trivial.

Meta-Reinforcement learning approaches aim to develop learning procedures that can adapt quickly to a distribution of tasks with the help of a few examples. Developing efficient exploration strategies capable of finding the most useful samples becomes critical in such settings. Existing approaches towards finding efficient exploration strategies add auxiliary objectives to promote exploration by the pre-update policy, however, this makes the adaptation using a few gradient steps difficult as the pre-update (exploration) and post-update (exploitation) policies are often quite different. Instead, we propose to explicitly model a separate exploration policy for the task distribution. Having two different policies gives more flexibility in training the exploration policy and also makes adaptation to any specific task easier. We show that using self-supervised or supervised learning objectives for adaptation allows for more efficient inner-loop updates and also demonstrate the superior performance of our model compared to prior works in this domain.

Computer-assisted synthesis planning aims to help chemists find better reaction pathways faster. Finding viable and short pathways from sugar molecules to value-added chemicals can be modeled as a retrosynthesis planning problem with a catalyst allowed. This is a crucial step in efficient biomass conversion. The traditional computational chemistry approach to identifying possible reaction pathways involves computing the reaction energies of hundreds of intermediates, which is a critical bottleneck in silico reaction discovery. Deep reinforcement learning has shown in other domains that a well-trained agent with little or no prior human knowledge can surpass human performance. While some effort has been made to adapt machine learning techniques to the retrosynthesis planning problem, value-added chemical discovery presents unique challenges. Specifically, the reaction can occur in several different sites in a molecule, a subtle case that has never been treated in previous works. With a more versatile formulation of the problem as a Markov decision process, we address the problem using deep reinforcement learning techniques and present promising preliminary results.