Oracle Imitation for Embedded Decision Making
Abstract
Major Goals: The project's major goal is to investigate a fundamentally new approach to decision making (planning) via imitation learning. The approach involves imprinting the behaviors of optimal solvers (oracles), such as Djikstra's algorithm, into neural networks to create an oracle network. The approach hypothesizes that by training a function approximator such as a neural network to mimic prior demonstrations from the oracle, we can learn embeddings of the problem and the oracles solutions and reproduce optimal solutions across both old and new problem instances while bypassing previous speed and computational complexity barriers associated with the producing behaviors using a neural embedding was motivated by looking at how Nature encodes intelligence. The human and animal brain generates behaviors in complex dynamic environments on instinct, rather than running internal tree-search algorithms on-the-fly. Our preliminary work has fundamentally shown oracle imitation as a computationally powerful and transformative approach to solving continuous domain planning problems. To expand this research and explore its potential to solve a wide variety of complex, computationally challenging planning problems, we will pursue the following are key research questions in this proposal: 1. What is an ideal architecture of oracle imitation to solve discrete, combinatorial problems, and how can it handle stochasticity in these systems? 2. How can oracle networks learn increasingly complex, hierarchical tasks that include nested discrete and continuous planning spaces? 3. How can oracle networks be trained to learn completely new tasks in new environments with as few queries to the oracle as possible?
Document Details
- Document Type
- Technical Report
- Publication Date
- May 16, 2022
- Accession Number
- AD1192848
Entities
People
- Michael Yip
Organizations
- University of California, San Diego