Task 2777.008: Learning Shields for Safe Multi-Agent Behavior from a Single Agent Policy

Event Date: February 14, 2019
Time: 2:00 pm EST/11 am PST
Priority: No
College Calendar: Show
Speaker: Joe Eappen, Purdue University

Abstract: Multi-agent learning is a considerable challenge faced by the Machine Intelligence community.  One of the problems that arises with training these systems is the difficulty for the agents to use previous knowledge learned from a single agent setting in a new environment with additional characteristics such as other agents.

In this talk we will look at a possible approach for transferring knowledge from a single agent policy to a multi-agent system by learning a shield (Alshiekh, 2018). This shielding function will be used to ensure safe behavior of the multi-agent system in a way that is close to the single agent objective. Preliminary results on a simple navigation task will be discussed as well as future directions of research.

Bio: Joe Kurian Eappen a first year graduate student in the School of Electrical and Computer Engineering at Purdue University working under the supervision of Prof. Suresh Jagannathan. He completed his Dual Degree (B.Tech & M.Tech) in Electrical Engineering at the Indian Institute of Technology (IIT) Madras, India in 2018. His current research is at the intersection of Reinforcement Learning, Multi-Agent systems and Verifiable Machine Learning.

 

Task 2777.008: Learning Shields for Safe Multi-Agent Behavior from a Single Agent Policy

Event Date: February 14, 2019
Time: 2:00 pm EST/11 am PST
Priority: No
College Calendar: Show
Speaker: Joe Eappen, Purdue University

Abstract: Multi-agent learning is a considerable challenge faced by the Machine Intelligence community.  One of the problems that arises with training these systems is the difficulty for the agents to use previous knowledge learned from a single agent setting in a new environment with additional characteristics such as other agents.

In this talk we will look at a possible approach for transferring knowledge from a single agent policy to a multi-agent system by learning a shield (Alshiekh, 2018). This shielding function will be used to ensure safe behavior of the multi-agent system in a way that is close to the single agent objective. Preliminary results on a simple navigation task will be discussed as well as future directions of research.

Bio: Joe Kurian Eappen a first year graduate student in the School of Electrical and Computer Engineering at Purdue University working under the supervision of Prof. Suresh Jagannathan. He completed his Dual Degree (B.Tech & M.Tech) in Electrical Engineering at the Indian Institute of Technology (IIT) Madras, India in 2018. His current research is at the intersection of Reinforcement Learning, Multi-Agent systems and Verifiable Machine Learning.