The success of deep reinforcement learning (DRL) hinges on the availability of training data, which is typically obtained via a large number of environment interactions. In many real-world scenarios, costs and risks are associated with gathering these data. The field of offline reinforcement learning addresses these issues through outsourcing the collection of data to a domain expert or a carefully monitored program and subsequently searching for a batch-constrained optimal policy. With the emergence of data markets, an alternative to constructing a dataset in-house is to purchase external data. However, while state-of-the-art offline reinforcement learning approaches have shown a lot of promise, they currently rely on carefully constructed datasets that are well aligned with the intended target domains. This raises questions regarding the transferability and robustness of an offline reinforcement learning agent trained on externally acquired data. In this paper, we empirically evaluate the ability of the current state-of-the-art offline reinforcement learning approaches to coping with the source-target domain mismatch within two MuJoCo environments, finding that current state-of-the-art offline reinforcement learning algorithms underperform in the target domain. To address this, we propose data valuation for offline reinforcement learning (DVORL), which allows us to identify relevant and high-quality transitions, improving the performance and transferability of policies learned by offline reinforcement learning algorithms. The results show that our method outperforms offline reinforcement learning baselines on two MuJoCo environments.
Over recent years, deep reinforcement learning has shown strong successes in complex single-agent tasks, and more recently this approach has also been applied to multi-agent domains. In this paper, we propose a novel approach, called MAGNet, to multi-agent reinforcement learning that utilizes a relevance graph representation of the environment obtained by a self-attention mechanism, and a message-generation technique. We applied our MAGnet approach to the synthetic predator-prey multi-agent environment and the Pommerman game and the results show that it significantly outperforms state-of-the-art MARL solutions, including Multi-agent Deep Q-Networks (MADQN), Multi-agent Deep Deterministic Policy Gradient (MADDPG), and QMIX
Learning to produce efficient movement behaviour for humanoid robots from scratch is a hard problem, as has been illustrated by the "Learning to run" competition at NIPS 2017. The goal of this competition was to train a two-legged model of a humanoid body to run in a simulated race course with maximum speed. All submissions took a tabula rasa approach to reinforcement learning (RL) and were able to produce relatively fast, but not optimal running behaviour. In this paper, we demonstrate how data from videos of human running (e.g. taken from YouTube) can be used to shape the reward of the humanoid learning agent to speed up the learning and produce a better result. Specifically, we are using the positions of key body parts at regular time intervals to define a potential function for potential-based reward shaping (PBRS). Since PBRS does not change the optimal policy, this approach allows the RL agent to overcome sub-optimalities in the human movements that are shown in the videos. We present experiments in which we combine selected techniques from the top ten approaches from the NIPS competition with further optimizations to create an high-performing agent as a baseline. We then demonstrate how video-based reward shaping improves the performance further, resulting in an RL agent that runs twice as fast as the baseline in 12 hours of training. We furthermore show that our approach can overcome sub-optimal running behaviour in videos, with the learned policy significantly outperforming that of the running agent from the video.
Robot navigation through crowds poses a difficult challenge to AI systems, since the methods should result in fast and efficient movement but at the same time are not allowed to compromise safety. Most approaches to date were focused on the combination of pathfinding algorithms with machine learning for pedestrian walking prediction. More recently, reinforcement learning techniques have been proposed in the research literature. In this paper, we perform a comparative evaluation of pathfinding/prediction and reinforcement learning approaches on a crowd movement dataset collected from surveillance videos taken at Grand Central Station in New York. The results demonstrate the strong superiority of state-of-the-art reinforcement learning approaches over pathfinding with state-of-the-art behaviour prediction techniques.
Microtubule networks (MTs) are a component of a cell that may indicate the presence of various chemical compounds and can be used to recognize properties such as treatment resistance. Therefore, the classification of MT images is of great relevance for cell diagnostics. Human experts find it particularly difficult to recognize the levels of chemical compound exposure of a cell. Improving the accuracy with automated techniques would have a significant impact on cell therapy. In this paper we present the application of Deep Learning to MT image classification and evaluate it on a large MT image dataset of animal cells with three degrees of exposure to a chemical agent. The results demonstrate that the learned deep network performs on par or better at the corresponding cell classification task than human experts. Specifically, we show that the task of recognizing different levels of chemical agent exposure can be handled significantly better by the neural network than by human experts.
Movement control of artificial limbs has made big advances in recent years. New sensor and control technology enhanced the functionality and usefulness of artificial limbs to the point that complex movements, such as grasping, can be performed to a limited extent. To date, the most successful results were achieved by applying recurrent neural networks (RNNs). However, in the domain of artificial hands, experiments so far were limited to non-mobile wrists, which significantly reduces the functionality of such prostheses. In this paper, for the first time, we present empirical results on gesture recognition with both mobile and non-mobile wrists. Furthermore, we demonstrate that recurrent neural networks with simple recurrent units (SRU) outperform regular RNNs in both cases in terms of gesture recognition accuracy, on data acquired by an arm band sensing electromagnetic signals from arm muscles (via surface electromyography or sEMG). Finally, we show that adding domain adaptation techniques to continuous gesture recognition with RNN improves the transfer ability between subjects, where a limb controller trained on data from one person is used for another person.
Curriculum Learning for Reinforcement Learning is an increasingly popular technique that involves training an agent on a defined sequence of intermediate tasks, called a Curriculum, to increase the agent's performance and learning speed. This paper introduces a novel paradigm for automatic curriculum generation based on a progression of task complexity. Different progression functions are introduced, including an autonomous online task progression based on the performance of the agent. The progression function also determines how long the agent should train on each intermediate task, which is an open problem in other task-based curriculum approaches. The benefits and wide applicability of our approach are shown by empirically comparing its performance to two state-of-the-art Curriculum Learning algorithms on a grid world and on a complex simulated navigation domain.
Deep RL approaches build much of their success on the ability of the deep neural network to generate useful internal representations. Nevertheless, they suffer from a high sample-complexity and starting with a good input representation can have a significant impact on the performance. In this paper, we exploit the fact that the underlying Markov decision process (MDP) represents a graph, which enables us to incorporate the topological information for effective state representation learning. Motivated by the recent success of node representations for several graph analytical tasks we specifically investigate the capability of node representation learning methods to effectively encode the topology of the underlying MDP in Deep RL. To this end we perform a comparative analysis of several models chosen from 4 different classes of representation learning algorithms for policy learning in grid-world navigation tasks, which are representative of a large class of RL problems. We find that all embedding methods outperform the commonly used matrix representation of grid-world environments in all of the studied cases. Moreoever, graph convolution based methods are outperformed by simpler random walk based methods and graph linear autoencoders.
Potential Based Reward Shaping combined with a potential function based on appropriately defined abstract knowledge has been shown to significantly improve learning speed in Reinforcement Learning. MultiGrid Reinforcement Learning (MRL) has further shown that such abstract knowledge in the form of a potential function can be learned almost solely from agent interaction with the environment. However, we show that MRL faces the problem of not extending well to work with Deep Learning. In this paper we extend and improve MRL to take advantage of modern Deep Learning algorithms such as Deep Q-Networks (DQN). We show that DQN augmented with our approach perform significantly better on continuous control tasks than its Vanilla counterpart and DQN augmented with MRL.
Real-world congestion problems (e.g. traffic congestion) are typically very complex and large-scale. Multiagent reinforcement learning (MARL) is a promising candidate for dealing with this emerging complexity by providing an autonomous and distributed solution to these problems. However, there are three limiting factors that affect the deployability of MARL approaches to congestion problems. These are learning time, scalability and decentralised coordination i.e. no communication between the learning agents. In this paper we introduce Resource Abstraction, an approach that addresses these challenges by allocating the available resources into abstract groups. This abstraction creates new reward functions that provide a more informative signal to the learning agents and aid the coordination amongst them. Experimental work is conducted on two benchmark domains from the literature, an abstract congestion problem and a realistic traffic congestion problem. The current state-of-the-art for solving multiagent congestion problems is a form of reward shaping called difference rewards. We show that the system using Resource Abstraction significantly improves the learning speed and scalability, and achieves the highest possible or near-highest joint performance/social welfare for both congestion problems in large-scale scenarios involving up to 1000 reinforcement learning agents.