Implement imitation learning framework for job-shop scheduling with DAgger and pilot heuristics

[Copilot]

Implements complete imitation learning framework for job-shop scheduling problems, supporting the LION20 conference paper on expert-like lookahead policies via pilot heuristics.

Core Package (pilotjsp/)

Instance Management (jsp_instance.py)

• OR-Library format I/O for 10×10 JSP instances
• Random instance generation with configurable seeds
• CSV export/import for data analysis

Feature Extraction (features.py)

• 13 dispatching rule features: PT, RPT, NRO, WR, ST, CR, SPT, EDD, FCFS, MW, MWT, ODD, TWC
• Batch extraction and normalization
• CSV persistence

Expert Solver (expert.py)

• Gurobi MIP formulation for optimal scheduling
• Disjunctive constraints for machine capacity
• Fallback to SPT heuristic when Gurobi unavailable

Preference Learning (preferences.py)

• Pairwise preferences from expert trajectories
• Rollout-based preference generation
• Training data aggregation with CSV I/O

DAgger Algorithm (dagger.py)

• Dataset aggregation with exponential beta decay
• Mixed expert/learned policy rollouts
• Training history tracking

Ordinal Regression (model.py)

• Hinge loss on preference pairs: max(0, -label × (score₁ - score₂) + margin)
• L2 regularization
• Feature importance via weight magnitudes

Pilot Heuristic (pilot.py)

• Configurable lookahead depth for decision evaluation
• Rollout simulation with learned base policy
• Makespan estimation for partial schedules

Utilities (utils.py)

• CSV/JSON serialization with numpy type handling
• Statistics computation
• Incremental logging

Configuration & Scripts

• config.yaml: Hyperparameters for expert, DAgger, model, pilot settings
• scripts/setup.py: Directory initialization and sample data
• scripts/run_example.py: End-to-end workflow demonstration
• scripts/evaluate.py: Baseline comparisons (SPT, LPT, FCFS, random)

Usage Example

from pilotjsp import JSPInstance, FeatureExtractor, OrdinalRegressionModel, PilotHeuristic

# Generate instance and extract features
instance = JSPInstance.generate_random(n_jobs=10, n_machines=10, seed=42)
extractor = FeatureExtractor(instance)

# Train model from preference pairs
model = OrdinalRegressionModel(n_features=13)
model.fit(features1, features2, labels)

# Evaluate pilot heuristic
pilot = PilotHeuristic(instance, extractor, model, rollout_depth=5)
solution = pilot.solve()
print(f"Makespan: {solution['makespan']}")

Dependencies

Requires: numpy, pyyaml, scikit-learn, matplotlib, pandas. Optional: gurobipy (for expert solver).

Original prompt

Create a Python repo “pilotJSP” for imitation learning in job-shop scheduling. Generate code to: make OR-Library 10x10 JSP instances, extract 13 DR features, query a Gurobi MIP expert, build preference pairs, run DAgger, train an ordinal-regression model, and evaluate a pilot rollout heuristic. Include CSV data handling, a config file, and basic scripts.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.