Bio-Robotics & Cybernetics
Machine Learning and Rock, Paper, Scissors
Using information learned both in the class and in the lab, this project, a group effort between three people, was accomplished with interesting results. The goal was to utilize machine learning to achieve a result of choice as long as it incorporated biological signals. This was accomplished in this particular case by using machine learning to create a system incapable of losing a game of rock, paper, scissors against a human opponent. In order to accomplish this, several types oof machine learning algorithms were employed including pipelines based on Random Forest algorithms, State Vector Machines, and K-means algorithms. These pipelines all included versions of signal transformations to eliminate noise from the EEG and EMG signals taken during data collection such as low pass filters, rectification, envelope detectors, and notch filters.
Results returned many interesting and varying bits of information. With no pipeline running raw data through the algorithm, the Random Forest algorithm performed near perfectly at about 91% accuracy however when added to a pipeline, classification accuracy dropped to near 39%. Looking through the code and stepping through the iterations, this was caused by an overclassification of available data caused by too small a data set. The pipeline chose data from the set more effectively eliminating the issue and showing a more accurate classification accuracy. However, combining a pipeline and an SVM allowed the accuracy to jump up to 62% accuracy utilizing the EEG and EMG signals. This was below the hopeful accuracy but showed that the algorithm was performing approximately twice as well as random guessing which can still be considered a success.