[cookiengineer]

The point behind a cyber defense system that is incorruptible is that it needs to find ways to represent learning mechanisms in an unsupervised manner. The only thing that is really unsupervised (as of state today) is an evolutionary concept, that is why I stuck with ES/HyperNEAT in the past. NEAT as a concept allows to implement flexible adapters for all kinds of things, while also allowing to have a time aware strategy planning neural network (the CPPN) if it's e.g. based on an unfolding LSTM layer(s). The reason why NEAT fits so nicely for this use case is the predictability of results of given tasks, and the possibility to change the neural layer structures based on the tasks and agents, too. So in our case we built e.g. LSTM layers for data structures that need time awareness, while e.g. a bayesian layer makes sense to represent an index map that refers to linked knowledge tree branches.

The previous implementation also featured a custom kademlia DHT which was implemented in a shrinkable and fast-forwardable manner so that the generation of "already outdated solutions to identical problems" could be just skipped ahead directly to the leaves of the knowledge tree to save redundant computation time.

And the only swarm intelligence that is well researched enough to represent a delegated learning mechanism based on bayesian/statistical truths are bee swarms [1] :)

In case you are interested in a project like this, that's basically what we are building @ tholian.network

[1] Check out Honeybee Democracy by Thomas Seeley