olegranmo
commented
4 years ago Hi @ricpruss! Started making documentation at readthedocs, where I will add some guidelines for setting the hyperparameters. In the meantime, basically, T decides how many clauses are involved in recognizing each particular pattern. If T=5, only five clauses are needed to represent each sub-pattern. A larger T value involves an increasing number of clauses, improving the representation of patterns. A rule of thumb seems to be that doubling both T and #clauses together automatically gives an increase in accuracy. The s-parameter decides the precision of the rules (frequency of the patterns). A larger s makes more detailed clauses, with lower frequency. For s, I typically explore values in the range 1.0 - 20.0. For T, exploring the range 1-200 may be worthwhile. However, with weighted clauses, much larger T-values must be used. With a max_weight of 255, multiplying T with 100 gives good results for classification, while multiplying with 10 seems to work well for regression.
ricpruss
s, T are defined in the paper but not given much explanation in the code or the paper really. For applying these models a little bit of guidance on effects of all the values would be super interesting.
class MultiClassTsetlinMachine(CommonTsetlinMachine): def init(self, number_of_clauses, T, s, boost_true_positive_feedback=1, number_of_state_bits=8, append_negated=True, max_weight=1, number_of_classes = None):