This repository contains all files to reproduce the results reported in the paper `Counterfactual Strategies for Markov Decision Processes' submitted to IJCAI 2025.
This repository was moved here.
├── requirements.txt # Requirements to re-run file
├── README.md
├── plots.ipynb # File to re-generate presented plots
├── benchmarks.py
├── journepy # Library for processing datasets.
├── data
│ ├── Store datasets here
├── out
│ ├── diversity
│ │ ├── Plots for Diversity
│ ├── models
│ │ ├── Stored model files
│ ├── user_strategies
│ | ├── Stored Strategy files
│ └── result files
├── Result.py
├── LogParser.py
├── gurobi_example.py
└── solver.py
Counterfactual strategies are generated by running the benchmarks.py script.
It offers several commandline arguments for specifying details of the execution:
-t, --timeout: Timeout for Gurobi
-s, --steps: Number of steps for each model
-i, --iterations: Iterations for each step
-c, --cores: Cores to use to parallelize experiments
-e, --experiments: Name of experiments to run
-rm, --rebuild_models: Rebuild models, implies rebuilding strategies
-rs, --rebuild_strategies: Rebuild strategies
-mi, --model_iterations: Number of models to generate for each setting
-as, --all_spotify: All spotify models in steps of 10% are generated
-d, --diversity_runs: Number of diverse counterfactuals
-b, --bounds: Bounds for gamma, evenly split by steps
Experiments can be from: 'greps', 'bpic12', 'bpic17-before', 'bpic17-after', 'bpic17-both', 'spotify', 'spotify1000', 'spotify2000', ... , 'spotify10000'.
The used datasets are not contained in the artifact due to their licensing.
However, all generated models and strategies are contained under out/models and out/user_strategies.
If you wish to generate models yourself, please download the respective datasets first to data/:
The artifact is written in Python and uses Gurobi.
Before running any code, please make sure that all requirements from requirements.txt are installed and that a valid Gurobi license is available.
To reproduce the results reported for GrepS, BPIC'12 and BPIC'17, run:
python3 benchmarks.py --experiments greps bpic12 bpic17-both --steps 10 --iterations 10 --model_iterations 1 --timeout 1800 --diversity_runs 2
To reproduce the detailed results reported for MSSD10 - MSSD30, run:
python3 benchmarks.py --experiments spotify1000 spotify2000 spotify3000 --steps 12 --iterations 10 --timeout 1800 --model_iterations 10 --diversity_runs 0 --bounds 0.1 0.5
To reproduce the general results for MSSD, run:
python3 benchmarks.py --experiments spotify --steps 10 --iterations 10 --model_iterations 10 --timeout 1800 --diversity_runs 0 --all_spotify
All presented plots can be reproduced by running the plots.ipynb notebook, the results file are contained in the repository.