Scholarlink

<p>Biological swarms are an invaluable source of inspiration in the design of artificial swarms. The complexity of collective behaviors often emerges from simpler individual behaviors. However, identifying simple models of complex behaviors involves ingenuity and non-trivial data analysis. We propose an approach to automate the process of extracting symbolic models from raw data gathered by observing a collective behavior. Our approach is composed of two phases. In the first, we train a graph neural network (GNN) to interpolate between raw data points and to isolate relationships across swarm members. In the second phase, we use the GNN embeddings to derive symbolic expressions (i.e., equations) through a two-layer evolutionary process which we call <i>macro-micro evolution</i> (MME). The top layer, <i>macro</i> evolution, manipulates the structure candidate equations; the bottom layer, <i>micro</i> evolution, tunes the parameters of a given equation. Experimental evaluation on both generated and real-world data shows that our approach produces compact expressions whose accuracy outperforms existing state-of-the-art techniques.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automatic extraction of symbolic models of collective behaviors with graph neural networks and macro-micro evolution

  • 期刊论文

  • 发表日期：

    2025年2月7日

查看全文
  • Stephen Powers,
  • Carlo Pinciroli

摘要

Biological swarms are an invaluable source of inspiration in the design of artificial swarms. The complexity of collective behaviors often emerges from simpler individual behaviors. However, identifying simple models of complex behaviors involves ingenuity and non-trivial data analysis. We propose an approach to automate the process of extracting symbolic models from raw data gathered by observing a collective behavior. Our approach is composed of two phases. In the first, we train a graph neural network (GNN) to interpolate between raw data points and to isolate relationships across swarm members. In the second phase, we use the GNN embeddings to derive symbolic expressions (i.e., equations) through a two-layer evolutionary process which we call macro-micro evolution (MME). The top layer, macro evolution, manipulates the structure candidate equations; the bottom layer, micro evolution, tunes the parameters of a given equation. Experimental evaluation on both generated and real-world data shows that our approach produces compact expressions whose accuracy outperforms existing state-of-the-art techniques.