Reinforcement Learning for Multi-Agent Competitive Scenarios

• Tipo: Artigo em Livro de Atas de Conferência Internacional
• Data: 2022

Autores

Coutinho, M (Autor) Outra; reis, lp (Autor) FEUP

Ata de Conferência Internacional

Título: 2022 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC)

Páginas: 130-135

Indexação

ISI Web of Knowledge - 0 Citações

Scopus - 0 Citações

Outras Informações

ID Authenticus: P-00W-ZNF

DOI: 10.1109/icarsc55462.2022.9784785

Abstract (EN): Reinforcement Learning techniques allow learning complex behaviors to deal with a variety of situations in a matter of hours. This complexity is even more prominent in multi-agent continuous 3D environments. This paper compares how the actions taken by two agents independently trained via a self-play approach differ from the ones taken when they are controlled by the same policy. It also explores the emergence of competitive or collaborative behaviors in a natural game setting. By implementing a 3D simulated version of the Dance Dance Revolution, the acquisition of more specific abilities like equilibrium, balance, and dexterity was tested. The approach achieved very good results learning a predefined sequence of buttons (7 arrows correctly pressed in 20M timesteps), revealing a similar learning behavior to human beings (improving with training and performing better in this kind of sequence than in random ones). The self-play approach also produced some interesting effects by developing cooperative behaviors in theoretically competitive scenarios.

Idioma: Inglês

Tipo (Avaliação Docente): Científica

Nº de páginas: 6