Enabling tool use and causal reasoning through combined symbolic and statistical machine learning

Abstract

We propose to construct a novel cognitive architecture that combines symbolic domain knowledge with the capability to acquire new skills using statistical machine learning. High level symbolic rules, provided by human input or by pre-programming, will place constraints on the transitions available to a graph-based statistical learning technique (such as a Markov decision process, MDP) using a technique we name transition matrix overlay construction, andexploration by the statistical learning system will generate new symbolic representations using a system based on a meta-critic.This architecture will be targeted specifically at spatial manipulation problems that involve causal reasoning, tool usage, and sequence learning, and is inspired by the capabilities of crows, ravens and other corvids. These birds are capable of solving a range of manipulation problems including using sticks to obtain out-of-reach food (tool usage), shaping wires into hooks to extract food from narrow crevices (tool modification), utilizing a short tool to obtain a longertool which can then be used to obtain food (meta-tool use or sequence learning) and learning to drop stones into a well to float food to the surface (causal reasoning). These skills are present in robots only to a very limited extent, but are grounded in realistic situations that commonly occur when operating in real-world environments. These problems require an integration of symbolicplanning approaches that can encapsulate information for future use with statistical methods that allow for real-time exploration and learning, while limiting the challenge of manipulation and perceptual tasks. We anticipate allowing robots to solve problems of similar complexity to those solved routinely by corvids. If successful, the flexibility and robustness that this system willgenerate would make it a fundamental addition to any robotic system that requires the capability to respond autonomously to variation in the environment or to generalize symbolic plans and knowledge into a robust, usable form for real-world operation.

Document Details

Document Type
DoD Grant Award
Publication Date
Sep 19, 2018
Source ID
N000141812776

Entities

People

  • Brian Scassellati

Organizations

  • Office of Naval Research
  • United States Navy
  • Yale University

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.
  • Robotics and Automation.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks
  • Autonomy