RL for Biology: Beyond Optimal Policies to Individual Differences

A new paper on arXiv (arXiv:2607.16542) titled "From Optimal Policies to Individual Differences: Rethinking Reinforcement Learning for Biology" tackles a fundamental blind spot in computational biology. Led by Patrick Govoni and five co-authors, the work critiques how RL is typically applied to model biological behavior. Standard RL algorithms optimize for a single best policy or an average across a population, but real biological systems exhibit persistent individual differences—animals of the same species behave differently even under identical conditions. The authors argue that this gap limits the explanatory power of RL-based models in fields like neuroscience, ecology, and evolutionary biology.

The paper explores promising approaches from various RL subfields that could bridge this gap. Techniques such as multi-task RL (learning diverse behaviors for different contexts), multi-objective RL (balancing conflicting goals), and population-based training (evolving a set of policies) offer ways to generate behavioral diversity. The authors emphasize the need for biologically plausible mechanisms—like noisy parameters, intrinsic motivation, or stochastic policies—that produce consistent, repeatable individuality rather than random noise. By integrating these ideas, future models could better capture how organisms explore, adapt, and differ from one another. This rethinking not only advances biological simulations but also inspires new RL architectures that embrace diversity over homogeneity.