Reinforcement Learning

Four Gymnasium-compatible RL environments for cancer immunotherapy optimization: checkpoint inhibitor dosing (anti-PD-1), combination dual checkpoint blockade (anti-PD-1 + anti-CTLA-4), CAR-T cell infusion scheduling, and adaptive dosing with pseudo-progression detection. Implements Kuznetsov-Taylor (1994) tumor-immune ODEs, Nikolopoulou (2018/2021) checkpoint inhibitor pharmacodynamics, Barros CARTmath (2021) CAR-T compartmental model, and Shulgin (2020) immune toxicity curves. Key finding: reward landscape curvature, not state dimensionality, determines RL difficulty - asymmetric drug toxicity profiles create richer gradient signals. 175 tests, MIT licensed.

Three Gymnasium-compatible RL environments for sepsis management: fluid resuscitation via bolus dosing, vasopressor titration for MAP maintenance, and combined multi-intervention management. Implements Reynolds et al. (2006) 4-ODE inflammation dynamics coupled with cardiovascular hemodynamics, lactate kinetics, and SOFA scoring. Three difficulty tiers per environment enable curriculum learning. PPO outperforms baselines on vasopressor titration with 10.6% improvement and lowest variance. 136 tests, MIT licensed.

Four Gymnasium-compatible RL environments for hemodialysis optimization: urea clearance via two-compartment Gotch-Sargent kinetics, ultrafiltration control with baroreceptor reflex cardiovascular dynamics, phosphate management with binder dosing across weekly cycles, and a full multi-objective dialysis session. Features three difficulty tiers, clinical protocol heuristic baselines, and PPO agents that exceed random baselines by up to 3.13x. 158 tests, MIT licensed.

Three Gymnasium-compatible RL environments for automated anesthesia drug dosing: propofol infusion control via the Marsh three-compartment pharmacokinetic model with Hill pharmacodynamic BIS prediction, remifentanil effect-site concentration targeting via the Minto model, and combined propofol-remifentanil anesthesia management using the Greco synergistic interaction surface. Includes configurable difficulty tiers with patient variability and surgical stimulation events, heuristic TCI clinical baselines, PPO RL agents, and a benchmark suite across three difficulty levels. Key finding: pharmacokinetic timescale, not task complexity, is the primary determinant of RL sample efficiency in drug dosing control. 109 tests, MIT licensed.

Gymnasium-compatible RL environment suite for brain-computer interfaces with three environments modeling motor imagery decoding, intracortical cursor control, and P300 speller navigation. Includes pluggable signal models (electrode drift, fatigue, co-adaptation, noise), a conditional VAE neural surrogate, CSP+LDA classical baseline, PPO RL baseline, and a five-tier benchmark suite. 158 tests, MIT licensed.

Open-source simulation platform providing a complete stack for training RL agents to navigate blood vessel graphs: TubeTK data ingestion, three Gymnasium environments with physics-based observations (VascularNav, FlowAwareNav, MagneticNav), analytical hemodynamics and magnetic field models, a neural flow surrogate, PPO baseline agents, and a benchmark suite across 5 difficulty tiers. 139 tests, MIT licensed.

Three Gymnasium-compatible RL environments for therapeutic peptide design: antimicrobial peptides (AMP), cyclic peptides, and T-cell epitopes. Includes heuristic biophysical scoring models, PPO and random baseline agents, reward shaping analysis revealing mode collapse boundaries, and a benchmark suite. First systematic demonstration that per-step reward shaping magnitude determines whether RL agents learn meaningful peptide sequences or degenerate to single-residue exploitation. 125 tests, MIT licensed.

Three Gymnasium-compatible RL environments for radiation therapy treatment planning: beam angle optimization, dose fractionation scheduling, and adaptive replanning. Includes analytical pencil beam dose calculation, linear-quadratic cell survival, TCP/NTCP radiobiological models, configurable difficulty tiers, and baseline agents (random, heuristic, PPO). PPO achieves 11.7x improvement on beam selection and 15.4x on adaptive replanning over clinical heuristics. 141 tests, MIT licensed.

Three Gymnasium-compatible RL environments for Type 1 diabetes glucose management: basal rate optimization, meal bolus dosing, and full closed-loop insulin delivery. Includes the Bergman minimal glucose-insulin model, Dalla Man gut absorption dynamics, a CGM sensor noise model, 30 virtual patients across three age groups, configurable difficulty tiers, heuristic clinical baselines, PPO RL agents, and a five-tier benchmark suite. Key finding: composite reward functions with safety constraints are necessary to differentiate learned policies from naive baselines in glucose management RL. 117 tests, MIT licensed.

Three Gymnasium-compatible RL environments for mechanical ventilation: tidal volume control via inspiratory pressure adjustment, PEEP optimization for oxygenation, and full ventilator parameter management for ARDS patients. Implements a single-compartment lung mechanics model coupled with a simplified gas exchange model, configurable difficulty tiers with patient variability and disease progression, heuristic clinical baselines, and PPO agents. Key finding: PPO improvement over baselines scales monotonically with action dimensionality, from 11.8% in 1D to 65.0% in 4D control. 230 tests, MIT licensed.

Three Gymnasium-compatible RL environments for cardiac electrophysiology: pacemaker rate optimization via the FitzHugh-Nagumo model with a cardiac conduction system simulator, antiarrhythmic drug dosing using the FitzHugh-Nagumo model with single-compartment PK/PD dynamics, and defibrillation timing via the Aliev-Panfilov model with probabilistic shock success. Includes configurable difficulty tiers, heuristic clinical baselines, and PPO agents. Environments span a difficulty spectrum from learnable (drug dosing, pacing) to open-challenge (defibrillation timing). 134 tests, MIT licensed.