Skip to content

AI Coevolution & Distributed EEDSE Engine (v3.2 Future Prospect)

Status: WIP / CIP Construction Site

This feature is strictly a Work-In-Progress (WIP) and Context-In-Process (CIP) construction site. AI is not used as a massive black box anywhere in PHIDS. Rather, AI-in-the-loop (AITL) is evaluated strictly as an interpretable assistant to Human-in-the-loop (HITL) exploration, ensuring researchers retain full biological oversight.

This document details the planned framework for distributed multi-objective Evolutionary Encapsulated Multi-Stage Design Space Exploration (EEDSE) and reinforcement learning-driven coevolutionary optimization in PHIDS.

1. Core Vision

While single-objective EEDSE (such as pymoo NSGA-III) successfully locates static Lotka-Volterra limit cycles, real-world ecosystems are driven by ongoing coevolutionary arms races. Flora species continuously adjust metabolic investment between morphological defenses (thorns) and induced volatile chemical signaling (VOCs), while herbivore species co-evolve specialized digestive efficiencies and chemical neutralization capabilities.

For detailed touchpoints on human vs. AI intervention gates across the pipeline, see DSE Governance & Interventions.

flowchart LR
    Flora_Pop["Flora Population<br><i>Defense Investment Strategy</i>"] <-->|Coevolutionary Feedback| Herbivore_Pop["Herbivore Population<br><i>Neutralization Strategy</i>"]

    SubGraph_Ray["Ray / Tune Distributed Cluster<br><i>Parallel Multi-Scenario Execution</i>"] --> Pareto["Pareto Optimal Front<br><i>Evolutionary Stable Strategies (ESS)</i>"]

2. Technical Architecture: Ray/Tune & Distributed Multi-Objective Optimization

  1. Distributed Cluster Parallelism: Utilizing Ray/Tune to scale scenario evaluations across HPC compute clusters (\(O(N_{\text{simulations}})\) concurrent workers).
  2. NSGA-III & Multi-Objective Pareto Optimization: Replaces single scalar cost functions with non-dominated sorting algorithms (NSGA-III via pymoo) targeting three concurrent objectives:
  3. Ecological Stability (\(S_{\text{LV}}\)): Spectral FFT limit cycle endurance.
  4. Empirical Distance (\(D_{\text{bio}}\)): Log-space Mahalanobis distance from empirical trait distributions (TRY/PanTHERIA).
  5. Defensive Diversity (\(H_{\text{chem}}\)): Shannon entropy of active plant secondary metabolites.
  6. Reinforcement Learning Agent Policies: Modeling herbivore swarms as MARL (Multi-Agent Reinforcement Learning) policies adapting foraging heuristics under dynamic plant defense induction.

3. Targeted Milestones

  • Phase 3.2.1: Ray/Tune task scheduler integration in phids.engine.batch.runner.
  • Phase 3.2.2: NSGA-III Pareto front extraction exporting balanced scenario blueprints directly to scenarios/*.yaml.
  • Phase 3.2.3: Dynamic SIMD bit-mask gene mutation passes on ECS trait structs during swarm mitosis and seed germination.