This is the pre-proceedings for the RLC 2026. You may expect minor changes.
Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
Will be presented at the Reinforcement Learning Conference (RLC), Montréal, Quebec, Canada, August 15–17, 2026.
Keywords: benchmark,gymnasium environment,nutritional regulation,reinforcement learning
We introduce NutriRL, a Gymnasium-compatible benchmark for homeostatic reinforcement learning under delayed state transitions, where the objective is to maintain cumulative levels of carbohydrates, fats, and proteins near a target nutritional balance over time. The environment models dietary regulation as a partially observable control problem in which each action produces delayed and gradual changes to an underlying nutritional state. These changes overlap and accumulate irreversibly, so rewards depend on state updates induced by past decisions across extended horizons. NutriRL is motivated by biological regulation in healthcare, where food intake produces delayed internal responses and effective diet control requires coordinating multiple macronutrients rather than matching aggregate caloric intake. Unlike common reinforcement learning formulations that reduce nutrition to scalar energy balancing or introduce delay only at the reward level, NutriRL embeds heterogeneous nutrient delays within state evolution itself. The agent must continuously regulate intake toward a hidden macronutrient target while avoiding overshoot, reflecting safety-sensitive control. Across representative value-based and policy-gradient methods, the benchmark reveals distinct sensitivity patterns as delay severity increases, along with a GRU-based recurrent policy to examine whether temporal memory mitigates delay-induced credit assignment challenges, illustrating how delayed transition dynamics influence learning stability and credit propagation. By isolating structural delay within a minimal, standardized, and computationally efficient model, NutriRL provides a reproducible testbed for evaluating temporal credit assignment under multi-channel delayed transition dynamics.
Aniket Khan, Charitha Palika, and V.Srinivasa Chakravarthy. "NutriRL: A Benchmark for Nutritional Regulation under Delayed State Transitions." Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
BibTeX:@article{khan2026nutrirl,
title={NutriRL: A Benchmark for Nutritional Regulation under Delayed State Transitions},
author={Aniket Khan and Charitha Palika and V.Srinivasa Chakravarthy},
journal={Reinforcement Learning Journal},
volume={7},
pages={},
year={2026}
}