Closed kymckay closed 2 years ago
Tried running the PSO using the absolute loss function (essentially just |y - y_hat| ) thinking it might simplify the search space, similar results to what we've been getting so far - all instances classified as the dominant class. The values of informants' best fitness appear to converge fairly early while the particles themselves appear to be all over the place even after 500 iterations.
https://www.sfu.ca/~ssurjano/spheref.html
Related to #16 - will show whether our implementation is working as expected and enable easier debugging