Offline Reinforcement Learning: A New Challenge for Symbolic Regression?

Created by W.Langdon from gp-bibliography.bib Revision:1.9101

@InProceedings{Heywood:2025:GPTP,

• author = "Bryce MacInnis and Malcolm Heywood",
• title = "Offline Reinforcement Learning: A New Challenge for Symbolic Regression?",
• booktitle = "Genetic Programming Theory and Practice XXII",
• year = "2025",
• editor = "Bogdan Burlacu and Fabricio {Olivetti de Franca} and Alexander Lalejini and Stephen Kelly and Wolfgang Banzhaf",
• series = "Genetic and Evolutionary Computation",
• pages = "189--210",
• address = "Michigan State University, USA",
• month = jun # " 5-7",
• publisher = "Springer",
• keywords = "genetic algorithms, genetic programming, MLP,ANN",
• isbn13 = "978-981-95-6397-5",
• code_url = "https://github.com/bmcnns/distributed-tpg",
• DOI = "10.1007/978-981-95-6398-2_10",
• abstract = "Offline reinforcement learning provides an approach to addressing the sample efficiency problem in reinforcement learning. For example, a behaviour policy could already have provided data for some costly (target) task. Additionally, reinforcement learning tasks that are described in terms of real-valued states and actions represent a particularly challenging scenario. With this in mind, we consider several offline reinforcement learning problems as a new challenge for symbolic regression. Our motivation being that genetic programming has had a long history of solving regression problems symbolically, i.e. interpretable solutions. Benchmarking two state-of-the-art symbolic regression methods on the Half Cheetah, Hopper and Walker2d locomotion tasks using offline data sourced from a deep reinforcement learning policy indicates that the resulting solutions appear to be limited by distributional shift (an incremental compounding of errors). Conversely, both a multi-layer perceptron and XGBoost are able to discover policies that replicate the performance of the original behavioural policy. The benchmarking datasets employed are publicly available for furthering research in this area.",
• notes = "published in 2026 after the workshop",

}

Genetic Programming entries for Bryce MacInnis Malcolm Heywood

Citations