This repository includes the code used in our work:

Sunger E., Bicer, Y., Erdogmus, Imbiriba, T., "MarkovType: A Markov Decision Process Strategy for Non-invasive BCI Systems", Proceedings of the AAAI Conference on Artificial Intelligence, 2025.

Please cite this paper (preprint) if you intend to use this code for your research.

This work proposes a Markov Decision Process for non-invasive BCI typing systems (MarkovType) and formulate the BCI typing procedure as a Partially Observable Markov Decision Process (POMDP), incorporating the typing mechanism into the learning procedure. We compare the performance of MarkovType with previous approaches using Recursive Bayesian Estimation following https://ieeexplore.ieee.org/document/10095715.

This repository was forked (then detached) from bci-disc-models, which is (c) 2022 Niklas Smedemark-Margulies and released under the MIT License.

We use https://pypi.org/project/thu-rsvp-dataset/1.1.0/ for fetching and preprocessing benchmark dataset from https://www.frontiersin.org/articles/10.3389/fnins.2020.568000/full.

Setup

Setup project with make and activate virtualenv with source venv/bin/activate

Usage

To reproduce our experiments, please follow these steps:

1. Preprocess data: python scripts/prepare_data.py
2. Pretrain baseline models: python scripts/train.py
3. Pretrain MarkovType models: python scripts/train_rnn.py
4. Evaluate models in simulated typing task: python scripts/evaluate.py
5. Parse saved results from evaluation with threshold: python scripts/parse_results.py
6. Parse saved results from evaluation without threshold: python scripts/parse_results_without_threshold.py
7. Collect statistics from parsed results: python scripts/analyze_results.py
8. Make plots: python scripts/plot_metrics.py