An Information-Theoretic Reinforcement Learning Framework for Autonomous Navigation
Abstract
This proposal will develop an enhanced Reinforcement Learning (RL)framework to tackle the problem of autonomous navigation in AUVs usingthe combination of a vision sensor that finds and recognizes objects in theworld, coupled with a centralized executive module that organizes the visualworld and associates cause-and-effect by interacting with the navigationsystem. The team is multidisciplinary and includes a cognitive scientist, amachine learning expert and Navy experts in AUV operations from theNaval Surface Warfare Center in Panama City. We opted for a two-prongapproach to include RL in AUV autonomy, coupling cognitive studies in RLlearning with mathematical innovation in RL modeling. Task 1 will definethe modules necessary to enable fast and efficient RL in a PARC forautonomous systems as well as to define the algorithms that govern theinteractions between those modules, as humans learn to navigateenvironments that contain threats and rewards. Task 2 will design a recurrentneural network that extracts events in time and organizes and stores them inassociative memory. Task 3 will include value of information in an RLmodel based framework to optimize the exploitation exploration dilemma.Task 4 will build a simulation environment appropriate for developing andtesting RL approaches, capable of generating sonar seafloor and targetimagery based on the simulated AUV~s actions within a world map whichcontains bathymetric, reflectivity, surface-type, and target data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 10, 2018
- Source ID
- N000141812306
Entities
People
- José PrÃncipe
Organizations
- Office of Naval Research
- United States Navy
- University of Florida