Adaptive Goal-Driven Autonomous Agents
Abstract
We propose research on autonomous agents endowed with multiple learning capabilities for robust performance under changing environme"ntal conditions. Specifically, our research focuses on the automated learning of knowledge for goal-driven agents. Goal-driven auton"omy (GDA) is a reflective model that controls the focus of an agent~s planning activities while monitoring the environment by checking for discrepancies. Discrepancies arise when the agent~s own expectations do not match its observations. Such discrepancies arise" when acting in partially observable environments that are also dynamic (i.e., changes occur independently from the agent~s actions)"". When discrepancies occur, the GDA agent formulates explanations for the discrepancies. Based on the explanation, the agent can cha""nge its goals or formulate alternative ones.We propose the automated learning of goal management, expectations, explanations, and"" goal formulation knowledge. Our proposed research combines two threads. First, we will represent GDA knowledge using both numerical" and symbolic information as opposed to purely symbolic representations as is the current practice. Reasoning on numerical quantitie"s is crucial for realistic applications where the agent requires explicit models of concepts such as time, costs, and ranges. Second"", we will use goal network formalisms to represent goal relationships beyond the traditional goal-subgoal associations. For learning"" goal management knowledge, we propose the automated learning of these goal network structures. Goal networks are hierarchical repre"sentations akin to HTNs but goals (instead of HTN~s tasks) are maintained at all layers of the hierarchy. For learning agent~s expec"tations, we propose a combination of symbolic and numerical regression techniques over the goal networks. For learning explanations" we propose to augment cause-effect relations with error margins and to combine a model for taxonomy s of expectations with probabil"istic truth maintenance propagated through the goal networks. For the task of learning goal formulation knowledge, we propose a Mark"ov-based model for prioritizing goals that combines numeric and symbolic information.This work will result in a new generation of" agents capable of operating under more realistic conditions. Crucially, our proposed research will result in a substantial reductio""n of the knowledge engineering effort required to deploy GDA technology in a variety of platforms, thus increasing the amount of aut"onomy available in a larger range of scenarios.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 06, 2017
- Source ID
- N000141812009
Entities
People
- Hector Munoz-avila
Organizations
- Lehigh University
- Office of Naval Research
- United States Navy