To main content

PINE: Planning and Identifying Neural Network for Thinking Fast and Slow

Abstract

The human brain's unique ability to seamlessly switch between quick, effortless reactions and slow, deliberate thinking forms the foundation for the content of this paper. Humans effortlessly navigate complex tasks like walking, speaking, and driving, relying on heuristic responses to their environment. This natural division of labor allows humans to adapt quickly and efficiently, tackling both new challenges and familiar tasks. Drawing inspiration from this cognitive duality, we propose, in this paper, a machine-based analog that employs a two-level functionality mirroring human capabilities. The approach centers on a Reinforcement Learning agent that replicates human learning through trial and error. The thesis introduces the Planning and Identifying Neural Network (PINE), a framework that identifies tasks using a combination of inherited environmental information, observations, and human-readable task descriptors. PINE leverages a model-based Reinforcement Learning algorithm, streamlining agent complexity and significantly reducing solution-finding time compared to model-free algorithms. PINE operates in real-time, utilizing initial environmental observations and task descriptors to identify and categorize encountered tasks. For unknown tasks, PINE employs meta-learning, selecting the most similar known task to expedite the solution-finding process with fewer episodes. Despite the challenge of task identification without interaction or labels, PINE demonstrates high accuracy in recognizing previously solved tasks and providing reliable estimates for the most similar tasks. This innovative two-level learning model represents a significant step towards creating adaptive machines that efficiently handle both familiar and novel challenges.

Category

Academic chapter

Language

English

Author(s)

Affiliation

  • SINTEF Digital / Sustainable Communication Technologies
  • University of Oslo

Year

2025

Publisher

IEEE (Institute of Electrical and Electronics Engineers)

Book

2025 11th International Conference on Robotics and Artificial Intelligence (ICRAI), December 19-21, 2025, Nagoya, Japan

ISBN

9798331590680

Page(s)

191 - 197

View this publication at Norwegian Research Information Repository