Running the tuxflow experiment, (sometimes) biasing the classifier output to boost flow with a Tux racer game.

How-to Tux2

Using Brain Products Liveamp, 32 electrodes.

Tested with OpenViBE 2.2.0 with Python 2.7 (windows & linux), R 3.4.4 through RKward (windows & linux).

Launch OpenViBE server, select liveamp driver, 32 electrodes, 500 Hz, with accelerometer. Set drift to 10ms. Use chunk size 4.

Setup

Runmi-csp-0-signal-monitoring.xml -- also possible check impedance through OpenViBE server.

Get Calibration data

Acquisition: launch tux, launch scenario mi-csp-1-acquisition.xml until end circuit. Two times.

Concatenation: run concatenate.R (cd to script directory first)

Select frequency: scenario mi-csp-1bis-frequencyBandSelection.xml

Cross-validation

NB: most of the code is takef from http://openvibe.inria.fr/tutorial-how-to-cross-validate-better/ ; see original scripts for more info. CSV Writer/Reader still use deprecated box because the new one at the moment necessitate a (useless here) signal stream, and do not let customize separators. Beware: match "cross_*" OpenViBE scenarios with their "regular" counterparts (e.g. filtering, epoching, etc.).

Requires "XML" library installed on R.

Run extract-labels.R, generate-folding.R, crossvalidate.R, aggregate.R.

Machine learning

Train spatial filter: mi-csp-2-train-CSP.xml

Train classifier: mi-csp-3-classifier-trainer.xml

Experiment

During XP: launch LSL2Joy

Between each run: generate new circuit.

For each run, launch Tux, launch mi-csp-4-online_noadapt.xml, start track special BCI.

Bonus

Participant can try regular track through scenario mi-csp-4bis-online_noadapt_truetrue.xml.

Files description

OpenViBE files

Taken from classical motor imagery scenarios, also contribution from Jussi Lindgren and Fabien Lotte.

• mi-csp-0-signal-monitoring: look at signals

• mi-csp-1-acquisition: record graz session

• mi-csp-1bis-frequencyBandSelection.xml: select band frequency

• mi-csp-2-train-CSP: train spatial filter

• mi-csp-3-classifier-trainer: train the (SVM) classifier

• mi-csp-4-online_noadapt: control tux / record data

• mi-csp-4-online_adapt: control tux / record data, this version bias the classifier with positive feedback

• replayer_adapt.xml: used in ./stats to retrieve online data

• cross_* files: mostly duplicate of regular scenarios, used for cross-validation on calibration

pyhton files used within scenarios

• python_lsl_stims.py: python box to retrieve stimulation from tux through LSL
• ov_bias_adapt_classifier_vidaurre.py: center output between run, bias depending of conditions, enable / disable classifier output in order to limit tux movements

Folders

• libs: required by the OpenViBE python box that fetch stimulation from tux racer.

NB: the file "liblsl.so" comes from a previous OpenViBE installation, for convenience. It may be needed to replace it with one coming from the computer running the expirement if OpenViBE crashes upon init (location: openvibe_folder/dependencies/lib).

• signals: empty on purpose, OpenViBE will write data in here, also conain most scenarios' output configuration to ease backup.

NB: two files will be recorded : one containing EEG data and the other the output of the classifier (*_EEG.gdf and *_class.gdf respectively). Both have the same stimulations (recorded from tux racer)

• cross_valid: used by scripts handling cross-validation on training sets in order to store temporary files

From tux1

• analyses: OpenViBE scenarios aimed at analysing offline previously recorded data
• data: mp3 for the "music" condition
• scripts: bash scripts that will facilitate the experiment (e.g. switch between the music)
• stats: first try for R statistics. Questionnaires data retrieved from the online formulaires, converted to CSV.