Skip to content

Latest commit

 

History

History
81 lines (57 loc) · 1.83 KB

QUICKSTART.md

File metadata and controls

81 lines (57 loc) · 1.83 KB

Quick Start Guide

1. Setup Environment

# Create virtual environment
python -m venv venv

# Activate (Windows PowerShell)
.\venv\Scripts\Activate.ps1

# Install dependencies
pip install -r requirements.txt

2. Verify Setup

python test_setup.py

This runs quick tests to ensure everything is installed correctly.

3. Train Agent

# Default training (2000 episodes)
python train.py

# Quick test (100 episodes)
python train.py --episodes 100

# Experiment with parameters
python train.py --episodes 3000 --cost-lambda 0.002 --gamma 0.98

4. View Results

After training, check:

  • outputs/plots/training_curve.png - Learning progress
  • outputs/plots/policy_rollouts.png - Recommended actions per state
  • outputs/plots/action_distribution.png - Action frequency analysis
  • outputs/validation_report.txt - Comprehensive statistics

5. Validate Saved Model

python train.py --validate-only --model-path outputs/models/best_model.pt

Expected Training Time

  • 100 episodes: ~1 minute
  • 2000 episodes: ~10-15 minutes (CPU)
  • 3000 episodes: ~20-30 minutes (CPU)

GPU training will be faster if CUDA is available.

Troubleshooting

Import errors

pip install --upgrade pip
pip install -r requirements.txt

YAML errors

Ensure config.yaml uses spaces (not tabs) for indentation.

Slow training

  • Reduce --episodes for testing
  • Use smaller batch_size in config.yaml
  • Enable CUDA if available: set device: "cuda" in config

Next Steps

  1. Sensitivity analysis: Test different cost_lambda values
  2. Longer horizons: Increase horizon_years in config
  3. State enrichment: Extend state space with age, traffic, etc.
  4. Portfolio: Adapt for multi-bridge scenarios
  5. Constraints: Add budget caps via action masking