Adaptive target localization under uncertainty using Multi-Agent Deep Reinforcement Learning with knowledge transfer

2024-11-01

Internet of Things

AI Summary

To address challenges in dynamic target localization within complex, unknown environments—including poor generalization and high cold-start overhead—this paper proposes a multi-agent deep reinforcement learning framework integrating knowledge transfer. Methodologically, it introduces, for the first time, a meta-learning-driven knowledge transfer module into a multi-agent Proximal Policy Optimization (PPO) architecture, augmented with graph neural network–based communication and uncertainty-aware modeling (entropy regularization and Bayesian reward estimation) to enable rapid cross-scenario policy adaptation. The core contribution is a generalizable, uncertainty-aware collaborative decision-making mechanism. Evaluated in both simulation and real-world drone swarm experiments, the framework achieves a 37% reduction in localization error, a 92.5% task completion rate, and a 4.8× improvement in sample efficiency—significantly enhancing system adaptability and training convergence speed.

Technology Category

Multiagent Systems: Adversarial AgentsSearch and Optimization: Learning to SearchMachine Learning: Transfer, Domain Adaptation, Multi-Task Learning

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Problem

• Multi-Robot Systems
• Unknown Target Localization
• Complex Environment

Innovation

• MADRL
• Knowledge Sharing
• Optimized Network Structure