kmdinake / somaco

SomAco is the algorithm I've developed for my COS 700 Research module.

SomAco

This repo presents a new algorithm that combines the meta-heuristics from ant colony optimization with the self-organizing feature map algorithm (SomAco) in the attempt of addressing the issue of slow training.

This algorithm was initially developed for my honours research module (COS700) at the University of Pretoria, under the supervision of Dr Christopher Cleghorn and Will van Heerden in the year 2018.

Developed by Keoagile M. Dinake

Copyright © 2019 Aistariot (Pty) Ltd.

Research Proposal

This browser does not support PDFs. Download the SomAco research proposal for offline usage.

somaco's Issues

New Stopping Condition

The algorithm tends to converge to a Lowest Std.Dev Training Error at around the 400th cycle. This leads me to wonder whether there should not be a 3rd stopping condition: one that will check for value convergence.

For instance, if the Lowest Std.Dev Training Error has been "constant" for the past arbitrarily chosen yet a reasonable number of epochs, then the Lowest Std.Dev Training Error has converged on a particular value as is unlikely to change even if the training goes beyond 1000 epochs (max number of epochs). As a result, the training should halt on that value.

Thus the new stopping condition is based on the following:

1. 0 <= Lowest Std.Dev Training Error <= 0.6 (hard constraint)
2. Epoch counter has reached the maximum number of epochs (soft constraint)
3. Lowest Std.Dev Training Error has reached convergence (hard constraint)

Early termination

The algorithm suffers from

early termination

due to the fact that the maximum number of epochs was insufficient