Modeling Goal Selection with Program Synthesis

J. Branson Byers, Bonan Zhao, Yael Niv

Track: Full track

Keywords: Reinforcement Learning, Program Inductions, Goals, Autonomous Agents

TL;DR: We formalize goal selection, and compare of model of goal selection with traditional reinforcement learning methods.

Abstract:

People can autonomously select and achieve novel goals to shape their own learning.
But goal selection can involve selecting goals from large spaces, where repeated
planning becomes computationally intractable. We propose program induction
as an inductive bias for defining human-like priors to make goal selection easier.
We demonstrate this tractable, semi-autonomous method for goal selection on a
novel ShapeWorld task using a handcrafted grammar that maps states to reward
functions.

Submission Number: 44