™
Hints for FLENN development
Speculations on what might help
2026 update
I'm still trying to figure out how to integrate fuzzy logic into neural networks in a productive way. My thoughts have evolved to bring into focus three aspects.First is reasoning: for example a planning system must be able to propose a variety of imaginable conditions, then develop the logical consequences of those conditions and also proposed responses. Is this an emergent mystery taking place somehow within the activation pattern of a NN system? Or is there a zone of construction with some kind of switching system, where therein various concepts available can be proposed and then the combination somehow used as input to a consequence-deriver to identify good and bad plans and outcomes. I imagine something like geometric space populated with imaginary agents and emotional valences at some level of abstraction and each able to interact and attract with others subject to some language or communication or resonance modality, and then a category or category-memory system connected to this sparking off proposals, then after settling or annealing or some computation one could say Great, that's got some of the parts of what we want, or No, scratch that let's start over, and gradually build out and ramify an effective plan or desireable path or something useful to the system, connected with the emotional valence that our inner direction-picker can use to decide what to do now about it. This is a sort of imaginasium, full of logic, categories, also emotional valuation. By connecting a rich representational system of category-avatars to it, we could really think about stuff.
Secondly, consider category formation, which we need to know how to do. And thirdly consider the benefits of leveled learning combined with a sort of sleep period in which the system can internally review recent experiences, in such a time the system would be able to carry out a vast search for new categories perhaps by applying a Fuzzy Logic combination-of-features learning algorithm to the very large number of possible input feature set combinations, through even standard back propagation learning. Level-based learning means that after learning to walk first, then we can learn the control categories and systems, then that's a level, only then we learn to run.
In such a system the acquisition/training of fuzzy logic category detectors can be done on the frontier of the current level, but then used in the trial-and-error phase in exploring and data collection for the next level. We need a metric of value for assessing the day's activities, maybe we can come up with one; then a review and search and reconsideration of combinations may reveal the best simplest control subsystem or category detecting discriminant feature.
This is my current uncertain thinking. If the learner moves from simpler to more complex tasks in a hierarchy of skill development, then each phase can produce categories which are useful in later phases. Concentration and authentic interest at one level supports all future activity because the categories there are strong.
Make sense?