Research Project ILP for AIBirds

This work is highly inspired by Renz and Zhang 2014, Qualitative Spatial Representation and Reasoning in Angry Birds: The Extended Rectangle Algebra.

Some parts for the execution and testing are from/inspired by https://github.com/logic-and-learning-lab/ilp-experiments

Reqirements

Only tested on linux
Python (3.10 used, lower and higher may work)
swipl
pip packages: pyswip, numpy, pandas, matplotlib, seaborn
Popper (Adjusted pip installation script to get it running as an executable, see logic-and-learning-lab/Popper#75)
optionally Aleph, installed using Swipl packages

Project Layout

aleph and popper folder: templates for learning
generate: Utilities for generating train and test data
data: generated train and test data
src: source files for background knowledge
solutions: resulting programs from popper, aleph and manual solutions for comparison
- sub folder intermediate: intermediate solutions to analyze the learn progression
evaluation: collated evaluation files from the solutionms on test data
scripts: Execution of generating, running and evaluating. Python scripts with options, bash scripts as complete execution

Available datasets

Sciencebirds: Randomly generated Scenes using the sciencebirds level generator (only regular rectangles)
Random: Randomly generated Scenes by dropping Objects from random points (additionally angular rectangles)

Additionally, each dataset has the option filtered. That means that for the supports predicate, that only those object combinations are valid, if the supporting object was removed. This woul roughly correspond to a supports_majority predicate

Examples

Learning the supports predicate with popper

python scripts/run_popper.py --target supports --dataset all

The solution will be saved to solutions/supports/popper_all

Evaluate the solution directly:

python scripts/test.py --set supports_all --dir test single solutions/supports/popper_all
# Optionally add --line to get results for each line
python scripts/test.py --set supports_all --dir test single solutions/supports/popper_all --line

Learning the supports predicate with aleph

python scripts/run_aleph.py --set all --target supports

The solution will be saved to solutions/supports/aleph_all

Evaluate equivalent to above

Batch Learning

Adjust scripts/run_experiments.sh to your liking and then run it.
Run scripts/evaluation.sh [target]. The plots are automatically generated using scripts/analysis.py

sargreal / bambirds-ilp