Attention-based communication for multi-agent reinforcement learning

Prelegent(ci)

Maciej Wojtala

Afiliacja

University of Warsaw

Język referatu

angielski

Termin

27 marca 2025 12:15

Pokój

p. 4050

Tytuł w języku polskim

Attention-based communication for multi-agent reinforcement learning

Seminarium

Seminarium „Gry, mechanizmy i sieci społeczne”

---

In multi-agent reinforcement learning, problems that attract most 
research include action-value function decomposition and inter-agent 
communication. In most studies, these problems are addressed separately. 
In this paper, we introduce the aggregation of messages from all agents 
in a central unit and the broadcast of the combined information to all 
the agents. Our proposed central unit can be easily added to existing 
architectures of action-value decomposition. Due to the use of attention 
in the central unit, it contains a number of trainable weights 
independent of the number of agents. In the experimental study with 
Foraging and SMAC2, our proposed approach yields encouraging results.

---