phil-hawkins / neural_heuristic_search

Neural heuristic Search

Neural Heuristic Search

Introduction

This repository is based on the paper: Modular Construction Planning using Graph Neural Network Heuristic Search

PyTorch and PyTorch Geometric are used to code the heuristic network.

Environment Setup

To create the environment with conda...

for CPU only

conda env create -f environment_cpu.yml

or with GPU support

conda env create -f environment_gpu.yml

Activate the environment

conda activate neural_heuristic_search

Running Examplar Searches

For the experiments in the paper we train against the results of A* searches for trusses spanning 1-3 struts/spokes

mkdir data
python build_training.py --target_dist 1 --train_example_path "./data/train_d1.pkl"
python build_training.py --target_dist 2 --train_example_path "./data/train_d2.pkl"
python build_training.py --target_dist 3 --train_example_path "./data/train_d3.pkl"

Training the Heuristic Network

Training the GIN based network:

python train.py --config GIN --epochs 2000 --batch_size 32 --lr 1e-3 --lr_patience 100 --train_data_search './data/train_d*.pkl'

Training the GAT based network:

python train.py --config GAT --epochs 2000 --batch_size 32 --lr 1e-3 --lr_patience 100 --train_data_search './data/train_d*.pkl'

Visualising Truss Construction Planning

The planning search and final construction plan can be viewed for a scenario

using a minimum Manhattan distance heuristic:

python scenario.py --target_dist 3 --heuristic Manhattan --show_search

using a mean Manhattan distance heuristic:

python scenario.py --target_dist 5 --heuristic Mean --show_search

using the GIN heuristic network:

python scenario.py --target_dist 5 --model_config GIN --show_search

Evaluation

Planning construction for 25 random scenarios with 5 span trusses:

python evaluate.py --planner_file demo_eval_planners.json --target_dist 5 --eps 1000 --max_scenarios 25 --batch_size 64 

Using demo_eval_planners.json this evaluates the following four agent heuristics for each scenario:

• GIN heuristic network
• GAT heuristic netork
• Minimum Manhattan distance heuristic
• Mean Manhattan distance heuristic

Results are output to ./logs/results.json by default