Tinashe2023/DecTwin-IPPO: DecTwin-PPO v1.0.0 — Code and data for "Fault-Aware Industrial Digital Twins via Cooperative Multi-Agent Reinforcement Learning" (EAAI submission)

Abstract

DecTwin-PPO v1.0.0 Reference implementation and evaluation data for the paper: "Fault-Aware Industrial Digital Twins via Cooperative Multi-Agent Reinforcement Learning: A Decentralized Partially Observable Markov Decision Process Architecture with Proximal Policy Optimization and Neighborhood-Aware Temporal Coordination" (submitted to Engineering Applications of Artificial Intelligence, October 2026). What's included Digital twin simulation environment: thermodynamic modeling (Newton's law of cooling), M/M/1 queue dynamics, Poisson fault injection, ring topologies of 4-16 machines IPPO agent with parameter sharing, GRU temporal processing, and neighbor-augmented observations (Dec-POMDP formulation) Rule-based failover baseline and centralized PPO baseline Training, inference, and figure-generation scripts Evaluation telemetry from all five experimental conditions reported in the paper (PostgreSQL export, CSV) Figures 1-6 as they appear in the paper Headline results (from the paper) Mean throughput 4.854 jobs/step: +145.5% vs rule-based baseline, +36.6% vs centralized PPO Zero-shot generalization across ring sizes 4-16 Graceful degradation at 3x the nominal fault rate Reproducing the paper See the README for setup. In brief: python scripts/train_marl.py (or use the included checkpoint) python scripts/inference_server.py + python scripts/baseline_runner.py python scripts/evaluate_marl.py regenerates Figures 3-6 from the telemetry data Citation If you use this code or data, please cite the paper (DOI will be added upon publication). Archived on Zenodo: [DOI placeholder] License MIT — see LICENSE.