Discussion items

5minAgenda itemAntti K.

Multi-agent research

In meeting

Antti K.: currently working on 3 projects:

  1. assist agents by observing only
  2. improving ad-hoc teammate performances by modeling other agents
  3. improving multiagents cooperation by optimizing beliefs over memory

The longer update focuses on (2). There a 2 agents and a set of items (several middle relevance, 1 huge relevance) to be collected from a gridworld. The 2 agents are fully cooperating. Currently the idea is to employ Meta Learning and CTDE (Centralized Training Decentralized Execution). It looks like ad-hoc cooperation provides better result with respect to the standard approach.

Action items