ECE Research Component of PhD Qualifier Exam by: Christian Ellis
When: Wednesday,
December 9, 2020
1:00 PM
-
3:00 PM
Where: > See description for location
Cost: Free
Description: Topic: A Path Towards Risk Averse Autonomous Vehicle Navigation
Zoom Teleconference: https://umassd.zoom.us/my/chellis
Abstract:
Current approaches to developing autonomous moving agents only allow the agent to operate in an environment representative of its previous experience. For agents that are modeled using a Markov decision process (MDP), a policy for navigation behavior is generated as the result of optimizing a reward function. Traditionally, this reward function is pre-defined by the algorithm designers and immutable. If the designed reward function fails to capture all aspects of the agent’s operational environment, undesired behavior will occur when the agent fails to optimize the true reward function and therefore express indifference to potentially dangerous, unseen scenarios. In the context of autonomous ground vehicles (AGV), consider an AGV which has been optimized in a wooded, off road environment described by a representative reward function. If the AGV is placed in a different environment such as an urban area, the original reward function will fail to accurately describe the desired behavior due to the presence of new terrain features, leading to potentially dangerous behavior. When an agent encounters features never seen before during training, how can an agent respond to these features? In these potentially dangerous scenarios (edge cases), an agent’s behavior should be risk adverse to decrease the chance of total system or mission failure.
This research explores the development of a risk adverse AGV by expressing uncertainty in the designed reward function by considering all possible reward functions that satisfy the training MDP. A Bayesian method is proposed to infer a reward function from a partially defined reward function based on human demonstrations and a training MDP, which enables risk averse behavior.
NOTE: All ECE Graduate Students are ENCOURAGED to join the zoom teleconference.
All interested parties are invited to join.
Advisor: Dr. Lance Fiondella
Committee Members: Dr. Liudong Xing and Dr. Hong Liu, Department of Electrical & Computer Engineering, University of Massachusetts Dartmouth
*For further information, please contact Dr. Lance Fiondella via email at lfiondella@umassd.edu.
Zoom Teleconference: https://umassd.zoom.us/my/chellis
Abstract:
Current approaches to developing autonomous moving agents only allow the agent to operate in an environment representative of its previous experience. For agents that are modeled using a Markov decision process (MDP), a policy for navigation behavior is generated as the result of optimizing a reward function. Traditionally, this reward function is pre-defined by the algorithm designers and immutable. If the designed reward function fails to capture all aspects of the agent’s operational environment, undesired behavior will occur when the agent fails to optimize the true reward function and therefore express indifference to potentially dangerous, unseen scenarios. In the context of autonomous ground vehicles (AGV), consider an AGV which has been optimized in a wooded, off road environment described by a representative reward function. If the AGV is placed in a different environment such as an urban area, the original reward function will fail to accurately describe the desired behavior due to the presence of new terrain features, leading to potentially dangerous behavior. When an agent encounters features never seen before during training, how can an agent respond to these features? In these potentially dangerous scenarios (edge cases), an agent’s behavior should be risk adverse to decrease the chance of total system or mission failure.
This research explores the development of a risk adverse AGV by expressing uncertainty in the designed reward function by considering all possible reward functions that satisfy the training MDP. A Bayesian method is proposed to infer a reward function from a partially defined reward function based on human demonstrations and a training MDP, which enables risk averse behavior.
NOTE: All ECE Graduate Students are ENCOURAGED to join the zoom teleconference.
All interested parties are invited to join.
Advisor: Dr. Lance Fiondella
Committee Members: Dr. Liudong Xing and Dr. Hong Liu, Department of Electrical & Computer Engineering, University of Massachusetts Dartmouth
*For further information, please contact Dr. Lance Fiondella via email at lfiondella@umassd.edu.
Contact:
ECE: Electrical & Computer Engineering Department 508.999.9164 http://www.umassd.edu/engineering/ece/
Topical Areas: General Public, University Community, College of Engineering, Electrical and Computer Engineering