A two-stage imitation learning framework for the multi-target search problem in swarm robotics

摘要

As a distributed system with a large number of individuals, swarm robotics is particularly suitable for multi-target search problems. Most existing work is about strategic design while this article focuses on strategy imitation. Sometimes we can observe the behavior of individuals and obtain a large amount of data, but we do not know the specific details of the strategy behind the behavior. Imitating the self-organizing behavior of organisms is of great significance for us to design efficient swarm strategies and to reveal the underlying mechanisms. The actual strategy adopted by individuals can be called the target strategy, and in this article, a two-stage imitation learning framework is proposed to approach the target strategy. In the first stage, a deep neural network is trained using the behavioral data of individuals, and in the second stage, the parameters of the neural network are further fine-tuned using the evolutionary algorithm. After two stages of learning and evolution, the resulting strategy RNSE is very close to the target strategy in terms of multiple indicators, including search efficiency, stability, parallel processing capability, and collaborative processing capability. In addition to multi-target search, the framework can also be used for other collective tasks such as aggregation and dispersion. In this paper, the design of neural networks and the settings of the evolutionary algorithm are discussed in detail, which is of great significance for the migration application of the framework. (C) 2019 Elsevier B.V. All rights reserved.