Article
Computer Science, Artificial Intelligence
Rex G. Liu, Michael J. Frank
Summary: A hallmark of human intelligence is the ability to compositionally generalize, which is challenging for reinforcement learning agents. This study proposes a hierarchical RL agent that learns and transfers task components and structures using a non-parametric Bayesian model, maintaining a factorized representation of task components through a hierarchical Dirichlet process.
ARTIFICIAL INTELLIGENCE
(2022)
Article
Robotics
Mukund Shah, Niravkumar Patel
Summary: This study introduces a flexible needle path planning algorithm that uses deep reinforcement learning methods to generate a kinematically feasible path for minimally invasive neurosurgery. The reinforcement learning algorithms are trained on segmented images of ventricles, blood vessels, and tumors, and the generated paths are compared with a traditional sampling-based algorithm. The results demonstrate that the proposed framework produces safer, shorter, and faster paths while avoiding critical structures.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2023)
Article
Engineering, Civil
Szilard Aradi
Summary: Academic research in the field of autonomous vehicles has gained popularity in recent years, covering various topics such as sensor technologies, communication, safety, decision making, and control. Artificial Intelligence and Machine Learning methods have become integral parts of this research. Motion planning, with a focus on strategic decision-making, trajectory planning, and control, has also been studied. This article specifically explores Deep Reinforcement Learning (DRL) as a field within Machine Learning. The paper provides insights into hierarchical motion planning and the basics of DRL, including environment modeling, state representation, perception models, reward mechanisms, and neural network implementation. It also discusses vehicle models, simulation possibilities, and computational requirements. The paper surveys state-of-the-art solutions, categorized by different tasks and levels of autonomous driving, such as car-following, lane-keeping, trajectory following, merging, and driving in dense traffic. Lastly, it raises open questions and future challenges.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2022)
Review
Engineering, Industrial
Ana Esteso, David Peidro, Josefa Mula, Manuel Diaz-Madronero
Summary: The objective of this paper is to examine the use and applications of reinforcement learning techniques in the production planning and control field. The analysis evaluated the characteristics, agent number and software tools of reinforcement learning, and reviewed 181 relevant articles. The results showed that reinforcement learning is mainly applied in production scheduling problems and has promising results, especially when considering uncertainty or non-linear properties.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
(2023)
Article
Multidisciplinary Sciences
Thomas McGrath, Andrei Kapishnikov, Nenad Tomasev, Adam Pearce, Martin Wattenberg, Demis Hassabis, Been Kim, Ulrich Paquet, Vladimir Kramnik
Summary: AlphaZero, a neural network engine that learns chess solely by playing against itself, acquires knowledge that enables it to outperform human chess players. Despite training without access to human games or guidance, it appears to learn concepts similar to those used by human chess players, as evidenced by linear probes and behavioral analysis.
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
(2022)
Article
Chemistry, Multidisciplinary
Tomoya Kawabe, Tatsushi Nishi, Ziang Liu
Summary: The use of multiple mobile robots has increased in logistics, manufacturing, and public services. Conflict-free route planning is a major research challenge for these robots. This article proposes a flexible route planning method that combines reinforcement learning and graph search algorithms for multiple robots.
APPLIED SCIENCES-BASEL
(2023)
Article
Computer Science, Artificial Intelligence
Yixing Lan, Xin Xu, Qiang Fang, Yujun Zeng, Xinwang Liu, Xianjian Zhang
Summary: This paper proposes a transfer reinforcement learning approach using auto-pruned decision trees for meta-knowledge extraction. Pre-trained policies are learned in source MDPs using RL algorithms, and meta-knowledge is extracted by re-training an auto-pruned decision tree model. In target MDPs, a hybrid policy integrating meta-knowledge and policies learned on the target MDPs is generated. Experimental results demonstrate that the proposed approach outperforms other baselines in terms of learning efficiency and interpretability.
KNOWLEDGE-BASED SYSTEMS
(2022)
Article
Computer Science, Artificial Intelligence
Lin Li, Wanzhong Zhao, Chunyan Wang, Abbas Fotouhi, Xuze Liu
Summary: In this study, an interactive merging strategy based on multi-agent deep reinforcement learning is proposed to coordinate the merging behaviors between ramp and mainline vehicles. By considering the dynamic reaction of mainline vehicles, the strategy successfully improves driving performance.
EXPERT SYSTEMS WITH APPLICATIONS
(2024)
Article
Automation & Control Systems
Junhyoung Ha, Byungchul An, Soonkyum Kim
Summary: In graph search algorithms, an environment is represented as a graph with system configurations and connections. A path connecting initial and goal configurations is generated by exploring the graph, aiming for optimality. Heuristic functions are used to guide the exploration efficiently. We propose RLHA*, a reinforcement learning approach using an artificial neural network as a heuristic function, to estimate optimal cost and achieve a suboptimal path. RLHA* outperforms existing methods in numerous simulations, demonstrating its consistent and robust performance.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
(2023)
Article
Construction & Building Technology
Sourav Dey, Thibault Marzullo, Gregor Henze
Summary: Reinforcement learning (RL) based control is a promising approach in building automation and control. However, it suffers from long training times and unstable control behavior during the early stages of its learning process, which makes it unsuitable to be applied directly to buildings. This paper addresses these issues using an inverse reinforcement learning approach (IRL), which learns the objective of a controller agent considered an expert in its domain.
ENERGY AND BUILDINGS
(2023)
Article
Computer Science, Artificial Intelligence
David Silver, Satinder Singh, Doina Precup, Richard S. Sutton
Summary: In this article, the hypothesis is made that intelligence and its associated abilities serve the maximization of reward, driving behavior that exhibits various abilities. It is suggested that agents learning through trial and error to maximize reward could exhibit most, if not all, of these abilities, potentially constituting a solution to artificial general intelligence.
ARTIFICIAL INTELLIGENCE
(2021)
Article
Neurosciences
Leonie Glitz, Keno Juechems, Christopher Summerfield, Neil Garrett
Summary: This study investigates the computational and neural mechanisms of information sharing and segmentation in planning. The findings suggest that human participants can efficiently switch between low-dimensional and high-dimensional representations of state transitions during planning tasks. The medial temporal lobe is identified as a key region for learning state transitions. The study also reveals that transition models are more strongly updated after positive outcomes.
JOURNAL OF NEUROSCIENCE
(2022)
Article
Computer Science, Artificial Intelligence
Zirui Xu, George P. Kontoudis, Kyriakos G. Vamvoudakis
Summary: We propose RRT-Q(8)(X), an online and intermittent kinodynamic motion planning framework for dynamic environments with unknown robot dynamics and unknown disturbances. The framework leverages RRTX for global path planning and rapid replanning to generate waypoints as a sequence of boundary-value problems (BVPs). We introduce a robust intermittent Q-learning controller for waypoint navigation with completely unknown system dynamics, external disturbances, and intermittent control updates. We demonstrate the effectiveness of RRT-Q(8)(X) through Monte Carlo numerical experiments in various dynamic and changing environments.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Telecommunications
Federico Venturini, Federico Mason, Francesco Pase, Federico Chiariotti, Alberto Testolin, Andrea Zanella, Michele Zorzi
Summary: The study proposed a distributed Reinforcement Learning approach that can scale to larger swarms without modifications. This method relies on UAVs exchanging information through a communication channel to achieve context-awareness and coordination of swarm actions. Experimental results show that the proposed method outperforms traditional look-ahead heuristic methods in handling non-uniform distributions of targets and obstacles.
IEEE TRANSACTIONS ON COGNITIVE COMMUNICATIONS AND NETWORKING
(2021)
Article
Computer Science, Artificial Intelligence
Alexandre Heuillet, Fabien Couthouis, Natalia Diaz-Rodriguez
Summary: The study explores the development of Explainable Reinforcement Learning (XRL) and the application of XAI techniques in helping to understand the behavior and internal workings of models in reinforcement learning. The evaluation focuses on studies directly linking explainability to RL, categorizing the explanation generation into transparent algorithms and post-hoc explainability. Furthermore, it reviews prominent XAI works and their potential impact on the latest advances in RL, addressing present and future everyday problems.
KNOWLEDGE-BASED SYSTEMS
(2021)
Article
Multidisciplinary Sciences
Johan Lind, Stefano Ghirlanda, Magnus Enquist
ROYAL SOCIETY OPEN SCIENCE
(2019)
Editorial Material
Psychology, Biological
Johan Lind
Summary: This article reports on a study by Wasserman, Kain, and O'Donoghue, which resolves the associative learning paradox by demonstrating that pigeons can solve complex category learning tasks through associative learning. The Outlook paper expands on this paradox and discusses the implications of their results.
LEARNING & BEHAVIOR
(2023)
Article
Multidisciplinary Sciences
Johan Lind, Vera Vinken, Markus Jonsson, Stefano Ghirlanda, Magnus Enquist
Summary: Identifying cognitive capacities underlying the human evolutionary transition is challenging. Recent studies suggest key differences in how humans and other animals recognize and remember information. This study tests the memory for stimulus sequences and finds that bonobos' working memory decays rapidly and they fail to learn the order of two stimuli, while humans solve the same sequence discrimination almost immediately. This indicates that non-human animals lack a memory for stimulus sequences, which may be one reason behind the origin of human culture.
Article
Biology
Michael T. J. Hague, H. Arthur Woods, Brandon S. Cooper
Summary: The widespread belief about a cognitive limit on human group size, known as 'Dunbar's number', is not supported by statistical analysis. Different methods yield wildly different numbers, with enormous confidence intervals making it futile to specify a single number for the cognitive limit.