Evolutionary Algorithms in the Design of BCI mice

One of the key features of our Analogue Evolutionary Brain-Computer Interfaces EPSRC Project (EP/F033818/1) was the use of evolutionary algorithms to aid the design of BCI mice. In particular, we used a technique known as Genetic Programming (GP), an evolutionary program-induction technology, in a variety of tasks (see further pointers on GP at the bottom of this page).

BCI mice are very sensitive to muscular movements, and in particular eye blinks. These are normally detected using Electro-oculogram (EOG) recorded by placing special electrodes on the subject's face. These electrodes, however, can easily detach and are not used very often. In a first piece of work we used GP to evolve algorithms that accurately approximate the behaviour of two standard detectors of ocular movement. The prediction was based entirely on EEG signals, i.e., without using EOG, making it possible to detect eye movements even in data recorded without EOG or eye tracking. Results of this method were excellent.

We also used GP as a means to synthesise complete brain-computer interfaces for mouse control. The objective was to evolve systems that analyse electroencephalographic signals and directly transform them into pointer movements, almost from scratch, the only input provided by us in the process being the set of visual stimuli to be used to generate recognisable brain activity. Experimental results with this approach were very promising and compared favourably with those produced by support vector machines.

In our analogue BCI mice we had traditionally used a simple technique for integrating the responses produced by the brain of subjects in the presence of the flashes of different stimuli (of which one was a target stimulus and several others were non-target stimuli). This technique worked well in controlled conditions. However, we found during the project that it was very sensitive to noise and particularly muscular artifacts. In a third piece of research we used GP to attack the problem of developing a more robust integration strategy capable of using the noisy and contradictory information provided at each time step by the signal processing systems into a coherent and precise trajectory for the mouse pointer. Results with the standard oddball stimulation sequences were extremely promising. Similarly promising results were obtained when we tested the approach on our new periodic stimulation sequences: again GP produced integration strategies that were much better than anything we could come up with ourselves at controlling the noise and artifacts present in brain signals.

Finally, we used GP in combination with our novel method for increasing the resolution of averages of event related potentials with excellent results.

Further Reading on GP

1. William B. Langdon and Riccardo Poli, Foundations of Genetic Programming, Springer, February 2002. (Second edition March 2005.) Cover of the Fundations of Genetic Programming book
2. Riccardo Poli, William B. Langdon and Nicholas Freitag McPhee, A Field Guide to Genetic Programming, with contributions by J.R. Koza, lulu.com, freely available under Creative Commons Licence, March 2008. Cover of Field Guide to Genetic Programming book
3. Riccardo Poli and Christopher R. Stephens, Taming the Complexity of Evolutionary Dynamics: From microscopic models to schema theory and beyond, Springer, Natural Computation series and Complexity series, approx 700 pages, forthcoming, 2012. Cover of Taming the Complexity of Evolutionary Dynamics book