For all its successes, Reinforcement Learning (RL) still struggles to deliver formal guarantees on the closed-loop behavior of the learned policy. Among other things, guaranteeing the safety of RL with respect to safety-critical systems is a very active research topic. Some recent contributions propose to rely on projections of the inputs delivered by the learned policy into a safe set, ensuring that the system safety is never jeopardized. Unfortunately, it is unclear whether this operation can be performed without disrupting the learning process. This paper addresses this issue. The problem is analysed in the context of $Q$-learning and policy gradient techniques. We show that the projection approach is generally disruptive in the context of $Q$-learning though a simple alternative solves the issue, while simple corrections can be used in the context of policy gradient methods in order to ensure that the policy gradients are unbiased. The proposed results extend to safe projections based on robust MPC techniques.

Interactive adaptive systems powered by Reinforcement Learning (RL) have many potential applications, such as intelligent tutoring systems. In such systems there is typically an external human system designer that is creating, monitoring and modifying the interactive adaptive system, trying to improve its performance on the target outcomes. In this paper we focus on algorithmic foundation of how to help the system designer choose the set of sensors or features to define the observation space used by reinforcement learning agent. We present an algorithm, value driven representation (VDR), that can iteratively and adaptively augment the observation space of a reinforcement learning agent so that is sufficient to capture a (near) optimal policy. To do so we introduce a new method to optimistically estimate the value of a policy using offline simulated Monte Carlo rollouts. We evaluate the performance of our approach on standard RL benchmarks with simulated humans and demonstrate significant improvement over prior baselines.

Multi-agent reinforcement learning (MARL) under partial observability has long been considered challenging, primarily due to the requirement for each agent to maintain a belief over all other agents' local histories -- a domain that generally grows exponentially over time. In this work, we investigate a partially observable MARL problem in which agents are cooperative. To enable the development of tractable algorithms, we introduce the concept of an information state embedding that serves to compress agents' histories. We quantify how the compression error influences the resulting value functions for decentralized control. Furthermore, we propose three natural embeddings, based on finite-memory truncation, principal component analysis, and recurrent neural networks. The output of these embeddings are then used as the information state, and can be fed into any MARL algorithm. The proposed embed-then-learn pipeline opens the black-box of existing MARL algorithms, allowing us to establish some theoretical guarantees (error bounds of value functions) while still achieving competitive performance with many end-to-end approaches.

This paper considers multi-agent reinforcement learning (MARL) in networked system control. Specifically, each agent learns a decentralized control policy based on local observations and messages from connected neighbors. We formulate such a networked MARL (NMARL) problem as a spatiotemporal Markov decision process and introduce a spatial discount factor to stabilize the training of each local agent. Further, we propose a new differentiable communication protocol, called NeurComm, to reduce information loss and non-stationarity in NMARL. Based on experiments in realistic NMARL scenarios of adaptive traffic signal control and cooperative adaptive cruise control, an appropriate spatial discount factor effectively enhances the learning curves of non-communicative MARL algorithms, while NeurComm outperforms existing communication protocols in both learning efficiency and control performance.

Emergent communication in artificial agents has been studied to understand language evolution, as well as to develop artificial systems that learn to communicate with humans. We show that agents performing a cooperative navigation task in various gridworld environments learn an interpretable communication protocol that enables them to efficiently, and in many cases, optimally, solve the task. An analysis of the agents' policies reveals that emergent signals spatially cluster the state space, with signals referring to specific locations and spatial directions such as "left", "up", or "upper left room". Using populations of agents, we show that the emergent protocol has basic compositional structure, thus exhibiting a core property of natural language.

Although in recent years reinforcement learning has become very popular the number of successful applications to different kinds of operations research problems is rather scarce. Reinforcement learning is based on the well-studied dynamic programming technique and thus also aims at finding the best stationary policy for a given Markov Decision Process, but in contrast does not require any model knowledge. The policy is assessed solely on consecutive states (or state-action pairs), which are observed while an agent explores the solution space. The contributions of this paper are manifold. First we provide deep theoretical insights to the widely applied standard discounted reinforcement learning framework, which give rise to the understanding of why these algorithms are inappropriate when permanently provided with non-zero rewards, such as costs or profit. Second, we establish a novel near-Blackwell-optimal reinforcement learning algorithm. In contrary to former method it assesses the average reward per step separately and thus prevents the incautious combination of different types of state values. Thereby, the Laurent Series expansion of the discounted state values forms the foundation for this development and also provides the connection between the two approaches. Finally, we prove the viability of our algorithm on a challenging problem set, which includes a well-studied M/M/1 admission control queuing system. In contrast to standard discounted reinforcement learning our algorithm infers the optimal policy on all tested problems. The insights are that in the operations research domain machine learning techniques have to be adapted and advanced to successfully apply these methods in our settings.

In this paper, we propose a model-free reinforcement learning method to synthesize control policies for motion planning problems for continuous states and actions. The robot is modelled as a labeled Markov decision process (MDP) with continuous state and action spaces. Linear temporal logics (LTL) are used to specify high-level tasks. We then train deep neural networks to approximate the value function and policy using an actor-critic reinforcement learning method. The LTL specification is converted into an annotated limit-deterministic B\"uchi automaton (LDBA) for continuously shaping the reward so that dense reward is available during training. A naive way of solving a motion planning problem with LTL specifications using reinforcement learning is to sample a trajectory and, if the trajectory satisfies the entire LTL formula then we assign a high reward for training. However, the sampling complexity needed to find such a trajectory is too high when we have a complex LTL formula for continuous state and action spaces. As a result, it is very unlikely that we get enough reward for training if all sample trajectories start from the initial state in the automata. In this paper, we propose a method that samples not only an initial state from the state space, but also an arbitrary state in the automata at the beginning of each training episode. We test our algorithm in simulation using a car-like robot and find out that our method can learn policies for different working configurations and LTL specifications successfully.

Behavioural cloning, where a computer is taught to perform a task based on demonstrations, has been successfully applied to various video games and robotics tasks, with and without reinforcement learning. This also includes end-to-end approaches, where a computer plays a video game like humans do: by looking at the image displayed on the screen, and sending keystrokes to the game. As a general approach to playing video games, this has many inviting properties: no need for specialized modifications to the game, no lengthy training sessions and the ability to re-use the same tools across different games. However, related work includes game-specific engineering to achieve the results. We take a step towards a general approach and study the general applicability of behavioural cloning on twelve video games, including six modern video games (published after 2010), by using human demonstrations as training data. Our results show that these agents cannot match humans in raw performance but can learn human-like behaviour. We also demonstrate how the quality of the data matters, and how recording data from humans is subject to a state-action mismatch, due to human reflexes.

Reinforcement learning (RL) has been successful in training agents in various learning environments, including video-games. However, such work modifies and shrinks the action space from the game's original. This is to avoid trying "pointless" actions and to ease the implementation. Currently, this is mostly done based on intuition, with little systematic research supporting the design decisions. In this work, we aim to gain insight on these action space modifications by conducting extensive experiments in video-game environments. Our results show how domain-specific removal of actions and discretization of continuous actions can be crucial for successful learning. With these insights, we hope to ease the use of RL in new environments, by clarifying what action-spaces are easy to learn.

Learning from Demonstration is increasingly used for transferring operator manipulation skills to robots. In practice, it is important to cater for limited data and imperfect human demonstrations, as well as underlying safety constraints. This paper presents a constrained-space optimization and reinforcement learning scheme for managing complex tasks. Through interactions within the constrained space, the reinforcement learning agent is trained to optimize the manipulation skills according to a defined reward function. After learning, the optimal policy is derived from the well-trained reinforcement learning agent, which is then implemented to guide the robot to conduct tasks that are similar to the experts' demonstrations. The effectiveness of the proposed method is verified with a robotic suturing task, demonstrating that the learned policy outperformed the experts' demonstrations in terms of the smoothness of the joint motion and end-effector trajectories, as well as the overall task completion time.