Search

The current reward learning from human preferences could be used to resolve complex reinforcement learning (RL) tasks without access to a reward function by defining a single fixed preference between pairs of trajectory segments. However, the judgment of preferences between trajectories is not dynamic and still requires human input over thousands of iterations. In this study, we proposed a weak human preference supervision framework, for which we developed a human preference scaling model that naturally reflects the human perception of the degree of weak choices between trajectories and established a human-demonstration estimator through supervised learning to generate the predicted preferences for reducing the number of human inputs. The proposed weak human preference supervision framework can effectively solve complex RL tasks and achieve higher cumulative rewards in simulated robot locomotion—MuJoCo games—relative to the single fixed human preferences. Furthermore, our established human-demonstration estimator requires human feedback only for less than 0.01% of the agent’s interactions with the environment and significantly reduces the cost of human inputs by up to 30% compared with the existing approaches. To present the flexibility of our approach, we released a video (https://youtu.be/jQPe1OILT0M) showing comparisons of the behaviors of agents trained on different types of human input. We believe that our naturally inspired human preferences with weakly supervised learning are beneficial for precise reward learning and can be applied to state-of-the-art RL systems, such as human-autonomy teaming systems. [ABSTRACT FROM AUTHOR]

The vehicle platoon will be the most dominant driving mode on future roads. To the best of our knowledge, few reinforcement learning (RL) algorithms have been applied in vehicle platoon control, which has large-scale action and state spaces. Some RL-based methods were applied to solve single-agent problems. If we need to tackle multiagent problems, we will use multiagent RL algorithms since the parameters space grows exponentially with the increasing number of agents involved. Previous multiagent RL algorithms generally may provide redundant information to agents, indicating a large amount of useless or unrelated information, which may cause to be difficult for convergence training and pattern extractions from shared information. Also, random actions usually contribute to crashes, especially at the beginning of training. In this study, a communication proximal policy optimization (CommPPO) algorithm was proposed to tackle the above issues. In specific, the CommPPO model adopts a parameter-sharing structure to allow the dynamic variation of agent numbers, which can well handle various platoon dynamics, including splitting and merging. The communication protocol of the CommPPO consists of two parts. In the state part, the widely used predecessor–leader follower typology in the platoon is adopted to transmit global and local state information to agents. In the reward part, a new reward communication channel is proposed to solve the spurious reward and “lazy agent” problems in some existing multiagent RLs. Moreover, a curriculum learning approach is adopted to reduce crashes and speed up training. To validate the proposed strategy for platoon control, two existing multiagent RLs and a traditional platoon control strategy were applied in the same scenarios for comparison. Results showed that the CommPPO algorithm gained more rewards and achieved the largest fuel consumption reduction (11.6%). [ABSTRACT FROM AUTHOR]

The unprecedented increase in data volume has become a severe challenge for conventional patterns of data mining and learning systems tasked with handling big data. The recently introduced Spark platform is a new processing method for big data analysis and related learning systems, which has attracted increasing attention from both the scientific community and industry. In this paper, we propose a shared nearest-neighbor quantum game-based attribute reduction (SNNQGAR) algorithm that incorporates the hierarchical coevolutionary Spark model. We first present a shared coevolutionary nearest-neighbor hierarchy with self-evolving compensation that considers the features of nearest-neighborhood attribute subsets and calculates the similarity between attribute subsets according to the shared neighbor information of attribute sample points. We then present a novel attribute weight tensor model to generate ranking vectors of attributes and apply them to balance the relative contributions of different neighborhood attribute subsets. To optimize the model, we propose an embedded quantum equilibrium game paradigm (QEGP) to ensure that noisy attributes do not degrade the big data reduction results. A combination of the hierarchical coevolutionary Spark model and an improved MapReduce framework is then constructed that it can better parallelize the SNNQGAR to efficiently determine the preferred reduction solutions of the distributed attribute subsets. The experimental comparisons demonstrate the superior performance of the SNNQGAR, which outperforms most of the state-of-the-art attribute reduction algorithms. Moreover, the results indicate that the SNNQGAR can be successfully applied to segment overlapping and interdependent fuzzy cerebral tissues, and it exhibits a stable and consistent segmentation performance for neonatal cerebral cortical surfaces. [ABSTRACT FROM AUTHOR]