Apprenticeship Learning for Heterogeneous Multi-agent Coordination
Author(s)
Xiong, Jerry Y.
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
We consider the problem of learning to coordinate a team of multiple, communicating agents without manually specifying a task-specific reward signal. In this thesis, we propose an approach that combines existing apprenticeship learning algorithms, which induce desirable behaviors by leveraging expert demonstrations, with an attention-based communication architecture. This approach enables sample-efficient learning of a high-level coordination strategy without manual specification of the low-level communication protocol required for such a strategy. We also introduce a novel mixed global/local discriminator architecture that is robust to heterogeneity in agents' observation and action spaces, along with several auxiliary optimization objectives to improve scalability for tasks with large numbers of agents. The proposed algorithm demonstrates effective learning on a variety of challenging tasks that require communication between heterogeneous agents.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis