Reinforcement learning as a comparative framework for the science of agency
Agency is attributed to a wide range of organisms across the tree of life. However, there is currently no systematic approach by which to investigate and compare agential phenomena and the diversity of agenthood. In this paper, I explore reinforcement learning (RL) as a candidate comparative framework for the science of agency. In Section 2, I introduce the history of RL and trace its roots in experimental research on animal learning before turning to its integration with computational methods and usage across disciplines. In Section 3, I consider what is needed of a cross-species framework for the scientific study of agency. In Section 4, I turn to RLs relationship to the concept of agency and outline how it can be used as an analytical tool to operationalise and model agential phenomena in Animalia and beyond. In Section 5, I synthesise these insights, assessing potential continuities and discontinuities that emerge from cross-species analysis. Once integrated with more naturalistic approaches, I claim that RL has broad theoretical scope as an empirical and formal framework for the science of agency I argue that comparative RL provides an indispensable (though partial) framework that accounts for the mechanics of agency without exhausting its organisation.
Authors
- Liberty Severs (ORCID: https://orcid.org/0000-0001-7961-6526)
Institutions
- University of Lisbon (PT)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24212201
- Primary Topic
- Philosophy of Biology and Scientific Explanation
- Type
- preprint