JFCM

Release 1.3 changelog:

• BaseConceptActivator.includePreviousOutput;
• change in output calculations: HyperbolicTangentActivator, LinearActivator, SigmoidActivator, SignumActivator include by default (includePreviousOutput = true) previous output in calculations;
• #17 Remove Jakarta Commons dependency, added StringUtils as substitute.

What has changed in calculations and algorithms, and why

Recently I’ve been in contact with Dr. Elpiniki I. Papageorgiou, I am so grateful to her because she gave me the opportunity to better understand theory of cognitive maps and learning algorithms, too! Dr. Papageorgiu is organizing a new book about fuzzy cognitive maps, and if everything goes well there will be a chapter about JFCM. Now that’s exciting! However, this means that the library calculations must be perfect, at least from an “academic” point of view, so I had to study a lot of material, especially regarding learning algorithms. After all this effort it was clear to me that the default calculations (until version 1.2-SNAPSHOT of JFCM), were not 100% accurate. The problem is that default ConceptActivator implementations used to calculate next output based only on incoming connections, completely ignoring previous value.

Now you have two possibilies:

• default implementations include previous value of Concept during calculation of Concept.nextOutput;
• you still can use the old algorithm, calling activator.setIncludePreviousOutput(false).

This change is going to break some applications, but the call above to setIncludePreviousOutput() should solve any compatibility issues.

Support material

In the Journals sections of Dr. Papageorgiou’s website you can find a lot of support material, the most important article being (for the above modifications):

E.I. Papageorgiou, C.D. Stylios & P.P. Groumpos, “Unsupervised learning techniques for fine-tuning Fuzzy Cognitive Map causal links”, Intern. Journal of Human-Computer Studies, Elsevier, vol. 64, 2006, pp. 727-743.