Deep reinforcement learning (DRL) is a promising method to learn control policies for robots only from demonstration and experience. To cover the whole dynamic behaviour of the robot, DRL training is an active exploration process typically performed in simulation environments. Although this simulation training is cheap and fast, applying DRL algorithms to real-world settings is difficult. If agents are trained until they perform safely in simulation, transferring them to physical systems is difficult due to the sim-to-real gap caused by the difference between the simulation dynamics and the physical robot. In this paper, we present a method of online training a DRL agent to drive autonomously on a physical vehicle by using a model-based safety supervisor. Our solution uses a supervisory system to check if the action selected by the agent is safe or unsafe and ensure that a safe action is always implemented on the vehicle. With this, we can bypass the sim-to-real problem while training the DRL algorithm safely, quickly, and efficiently. We compare our method with conventional learning in simulation and on a physical vehicle. We provide a variety of real-world experiments where we train online a small-scale vehicle to drive autonomously with no prior simulation training. The evaluation results show that our method trains agents with improved sample efficiency while never crashing, and the trained agents demonstrate better driving performance than those trained in simulation.

We study the adaption of soft actor-critic (SAC) from continuous action space to discrete action space. We revisit vanilla SAC and provide an in-depth understanding of its Q value underestimation and performance instability issues when applied to discrete settings. We thereby propose entropy-penalty and double average Q-learning with Q-clip to address these issues. Extensive experiments on typical benchmarks with discrete action space, including Atari games and a large-scale MOBA game, show the efficacy of our proposed method.

Bayesian policy reuse (BPR) is a general policy transfer framework for selecting a source policy from an offline library by inferring the task belief based on some observation signals and a trained observation model. In this paper, we propose an improved BPR method to achieve more efficient policy transfer in deep reinforcement learning (DRL). First, most BPR algorithms use the episodic return as the observation signal that contains limited information and cannot be obtained until the end of an episode. Instead, we employ the state transition sample, which is informative and instantaneous, as the observation signal for faster and more accurate task inference. Second, BPR algorithms usually require numerous samples to estimate the probability distribution of the tabular-based observation model, which may be expensive and even infeasible to learn and maintain, especially when using the state transition sample as the signal. Hence, we propose a scalable observation model based on fitting state transition functions of source tasks from only a small number of samples, which can generalize to any signals observed in the target task. Moreover, we extend the offline-mode BPR to the continual learning setting by expanding the scalable observation model in a plug-and-play fashion, which can avoid negative transfer when faced with new unknown tasks. Experimental results show that our method can consistently facilitate faster and more efficient policy transfer.

Due to the vastly different energy consumption between up-slope and down-slope, a path with the shortest length on a complex off-road terrain environment (2.5D map) is not always the path with the least energy consumption. For any energy-sensitive vehicle, realizing a good trade-off between distance and energy consumption in 2.5D path planning is significantly meaningful. In this paper, we propose a deep reinforcement learning-based 2.5D multi-objective path planning method (DMOP). The DMOP can efficiently find the desired path in three steps: (1) Transform the high-resolution 2.5D map into a small-size map. (2) Use a trained deep Q network (DQN) to find the desired path on the small-size map. (3) Build the planned path to the original high-resolution map using a path-enhanced method. In addition, the hybrid exploration strategy and reward shaping theory are applied to train the DQN. The reward function is constructed with the information of terrain, distance, and border. Simulation results show that the proposed method can finish the multi-objective 2.5D path planning task with significantly high efficiency. With similar planned paths, the speed of the proposed method is more than 100 times faster than that of the A* method and 30 times faster than that of H3DM method. Also, simulation proves that the method has powerful reasoning capability that enables it to perform arbitrary untrained planning tasks.

Prescriptive business process monitoring provides decision support to process managers on when and how to adapt an ongoing business process to prevent or mitigate an undesired process outcome. We focus on the problem of automatically reconciling the trade-off between prediction accuracy and prediction earliness in determining when to adapt. Adaptations should happen sufficiently early to provide enough lead time for the adaptation to become effective. However, earlier predictions are typically less accurate than later predictions. This means that acting on less accurate predictions may lead to unnecessary adaptations or missed adaptations. Different approaches were presented in the literature to reconcile the trade-off between prediction accuracy and earliness. So far, these approaches were compared with different baselines, and evaluated using different data sets or even confidential data sets. This limits the comparability and replicability of the approaches and makes it difficult to choose a concrete approach in practice. We perform a comparative evaluation of the main alternative approaches for reconciling the trade-off between prediction accuracy and earliness. Using four public real-world event log data sets and two types of prediction models, we assess and compare the cost savings of these approaches. The experimental results indicate which criteria affect the effectiveness of an approach and help us state initial recommendations for the selection of a concrete approach in practice.

Transformers have significantly impacted domains like natural language processing, computer vision, and robotics, where they improve performance compared to other neural networks. This survey explores how transformers are used in reinforcement learning (RL), where they are seen as a promising solution for addressing challenges such as unstable training, credit assignment, lack of interpretability, and partial observability. We begin by providing a brief domain overview of RL, followed by a discussion on the challenges of classical RL algorithms. Next, we delve into the properties of the transformer and its variants and discuss the characteristics that make them well-suited to address the challenges inherent in RL. We examine the application of transformers to various aspects of RL, including representation learning, transition and reward function modeling, and policy optimization. We also discuss recent research that aims to enhance the interpretability and efficiency of transformers in RL, using visualization techniques and efficient training strategies. Often, the transformer architecture must be tailored to the specific needs of a given application. We present a broad overview of how transformers have been adapted for several applications, including robotics, medicine, language modeling, cloud computing, and combinatorial optimization. We conclude by discussing the limitations of using transformers in RL and assess their potential for catalyzing future breakthroughs in this field.

We present MDP Playground, a testbed for Reinforcement Learning (RL) agents with dimensions of hardness that can be controlled independently to challenge agents in different ways and obtain varying degrees of hardness in toy and complex RL environments. We consider and allow control over a wide variety of dimensions, including delayed rewards, sequence lengths, reward density, stochasticity, image representations, irrelevant features, time unit, action range and more. We define a parameterised collection of fast-to-run toy environments in OpenAI Gym by varying these dimensions and propose to use these to understand agents better. We then show how to design experiments using MDP Playground to gain insights on the toy environments. We also provide wrappers that can inject many of these dimensions into any Gym environment. We experiment with these wrappers on Atari and Mujoco to allow for understanding the effects of these dimensions on environments that are more complex than the toy environments. We also compare the effect of the dimensions on the toy and complex environments. Finally, we show how to use MDP Playground to debug agents, to study the interaction of multiple dimensions and describe further use-cases.

World models power some of the most efficient reinforcement learning algorithms. In this work, we showcase that they can be harnessed for continual learning - a situation when the agent faces changing environments. World models typically employ a replay buffer for training, which can be naturally extended to continual learning. We systematically study how different selective experience replay methods affect performance, forgetting, and transfer. We also provide recommendations regarding various modeling options for using world models. The best set of choices is called Continual-Dreamer, it is task-agnostic and utilizes the world model for continual exploration. Continual-Dreamer is sample efficient and outperforms state-of-the-art task-agnostic continual reinforcement learning methods on Minigrid and Minihack benchmarks.

The main challenge of offline reinforcement learning, where data is limited, arises from a sequence of counterfactual reasoning dilemmas within the realm of potential actions: What if we were to choose a different course of action? These circumstances frequently give rise to extrapolation errors, which tend to accumulate exponentially with the problem horizon. Hence, it becomes crucial to acknowledge that not all decision steps are equally important to the final outcome, and to budget the number of counterfactual decisions a policy make in order to control the extrapolation. Contrary to existing approaches that use regularization on either the policy or value function, we propose an approach to explicitly bound the amount of out-of-distribution actions during training. Specifically, our method utilizes dynamic programming to decide where to extrapolate and where not to, with an upper bound on the decisions different from behavior policy. It balances between the potential for improvement from taking out-of-distribution actions and the risk of making errors due to extrapolation. Theoretically, we justify our method by the constrained optimality of the fixed point solution to our $Q$ updating rules. Empirically, we show that the overall performance of our method is better than the state-of-the-art offline RL methods on tasks in the widely-used D4RL benchmarks.

In the intrinsically motivated skills acquisition problem, the agent is set in an environment without any pre-defined goals and needs to acquire an open-ended repertoire of skills. To do so the agent needs to be autotelic (deriving from the Greek auto (self) and telos (end goal)): it needs to generate goals and learn to achieve them following its own intrinsic motivation rather than external supervision. Autotelic agents have so far been considered in isolation. But many applications of open-ended learning entail groups of agents. Multi-agent environments pose an additional challenge for autotelic agents: to discover and master goals that require cooperation agents must pursue them simultaneously, but they have low chances of doing so if they sample them independently. In this work, we propose a new learning paradigm for modeling such settings, the Decentralized Intrinsically Motivated Skills Acquisition Problem (Dec-IMSAP), and employ it to solve cooperative navigation tasks. First, we show that agents setting their goals independently fail to master the full diversity of goals. Then, we show that a sufficient condition for achieving this is to ensure that a group aligns its goals, i.e., the agents pursue the same cooperative goal. Our empirical analysis shows that alignment enables specialization, an efficient strategy for cooperation. Finally, we introduce the Goal-coordination game, a fully-decentralized emergent communication algorithm, where goal alignment emerges from the maximization of individual rewards in multi-goal cooperative environments and show that it is able to reach equal performance to a centralized training baseline that guarantees aligned goals. To our knowledge, this is the first contribution addressing the problem of intrinsically motivated multi-agent goal exploration in a decentralized training paradigm.