Paper Image

Planning with imagination: Enabling neural networks to perform high-level reasoning in learned abstract search spaces

Published on:

16 August 2023

Primary Category:

Artificial Intelligence

Paper Authors:

Carlos Martin,

Tuomas Sandholm

Bullets

Key Details

Proposes a novel reinforcement learning architecture for planning in learned abstract spaces

Enables reasoning via compound, temporally-extended actions unlike real environment steps

Planning space is fully decoupled from real environment dynamics

Handles continuous actions and partial observability

Outperforms prior methods in multiple benchmark domains

AI generated summary

Planning with imagination: Enabling neural networks to perform high-level reasoning in learned abstract search spaces

This paper proposes a new reinforcement learning method that gives agents the ability to perform high-level planning by imagining and searching through a learned abstract space. The key advantage is that planning can occur at arbitrary timescales, enabling reasoning in terms of compound actions rather than individual environment steps. This approach fully decouples the imagined planning space from the real environment's dynamics. Experiments on navigation, puzzle, and optimization tasks demonstrate superior performance over comparable methods.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up