Key points are not available for this paper at this time.
Genetic programming (GP) is a popular problem-solving and optimization technique. However, generating effective test cases for training and evaluating GP programs requires strong domain knowledge. Furthermore, GP programs often prematurely converge on local optima when given excessively difficult problems early in their training. Curriculum learning (CL) has been effective in addressing similar issues across different reinforcement learning (RL) domains, but it requires the manual generation of progressively difficult test cases as well as their careful scheduling. In this work, we leverage the domain knowledge and the strong generative abilities of large language models (LLMs) to generate effective test cases of increasing difficulties and schedule them according to various curricula. We show that by integrating a curriculum scheduler with LLM-generated test cases we can effectively train a GP agent player with environments-based curricula for a single-player game and opponent-based curricula for a multi-player game. Finally, we discuss the benefits and challenges of implementing this method for other problem domains.
Building similarity graph...
Analyzing shared references across papers
Loading...
Steven Jorgensen
Moscow Institute of Thermal Technology
Giorgia Nadizar
Centre National de la Recherche Scientifique
Gloria Pietropolli
Centrum Wiskunde & Informatica
University of Trieste
IIT@MIT
MIT Lincoln Laboratory
Building similarity graph...
Analyzing shared references across papers
Loading...
Jorgensen et al. (Mon,) studied this question.
synapsesocial.com/papers/6a10d2858102eb4b66ee7ca9 — DOI: https://doi.org/10.1145/3638529.3654056