Focused Macros

Learn focused macros (action abstractions) for efficient black-box planning

Init Setup

git clone --recursive [email protected]:xyntechx/focused-macros.git
cd focused-macros
python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip wheel
pip install -r requirements.txt

Running

Activate virtual environment

source .venv/bin/activate

Update submodules

git submodule update --recursive

Learn focused macros

python3 macro_learning.py

Run planning

python3 planning.py

Visualize focused macros usage

python3 visualize.py

Solve cube with a saved plan

python3 solve.py

Results

Macro Learning

The learned focused macros which enable the shortest planning times are saved in output/n10_learned_macros.pkl. Here are the specs:

Max Fringe Size: N_m * 10
Learning Time: 06:38
# of Macros Generated: 576
Best Net Effect (h) Heuristic: 6
Worst Net Effect (h) Heuristic: 16
Shortest Macro Length (g): 6
Longest Macro Length (g): 16

List of learned focused macros .pkl files in output:

• n10_learned_macros.pkl: learned through greedy BFS with a fringe of max size N_m * 10
• inf_learned_macros.pkl: learned through greedy BFS with an unlimited fringe

Planning

Plans are saved in output/plans.

Data from planning is saved in output/data. Each .pkl file is a pandas df storing the effect size and length of each focused macro (or primitive action) used at each timestep of the plan.

Graphs of effect size against timestep and macro length against timestep are saved in output/plots.

For each of these folders, its subfolders correspond to the max fringe size used for macro learning (e.g. n10 means the focused macros used were learned through greedy BFS with a fringe of max size N_m * 10).