OpenKAGE: Open-ended lifelong learning in the open world via knowledge-sharing agent-group evolution
OpenKAGE: Open-Ended Lifelong Learning in the Open World via Knowledge-Sharing Agent-Group Evolution
The authors propose OpenKAGE, an agent-group evolution framework for open-ended lifelong learning in the open world, enabling agents to continuously acquire new abilities by accumulating and reusing their own and others' experience. The framework has two layers of knowledge representation: a memory layer storing episodic memory and an experience layer storing transferable experience and skills along with the conditions for reuse; a knowledge-sharing mechanism semantically integrates textual experience into a continuously evolving shared knowledge graph, while executable skills are shared as-is.
The paper presents a two-layer knowledge representation and a shared knowledge graph mechanism, offering insight into how agent groups keep learning and transfer across models in open worlds.
Source: arXiv Game AI · arxiv.org