A system makes a decision basing on a large number of varied factors, following a "live" decision tree - one that is (independently, through other subsystem) updated with new decisions, new situations.
The individual decisions can be recorded as a kind of structure:
- decision function
- node to activate if decision is positive
- node to activate if decision is negative
and a node can be another decision record, or a conclusion.
This isn't entirely a binary tree, as many decisions may lead to the same conclusion - each node has two children, but may have many parents.
There is absolutely no problem storing the tree in memory - it can be database records or entries of a map, or just a list. It's perfectly sufficient for the machine.
The problem here is building the subsystem that expands the decision tree - and in particular, having a human operator understand the structure being built, to be able to tune, guide, fix, adjust it: debugging the AI learning process.
The question is: how to represent that data in a human-readable way, that emphasizes the flow of the graph?
a non-working example of the answer is Concept map - in this case it only goes so far; with more than thirty or so nodes, it becomes a jumbled mess, especially if the number of cross-connections (multiple parents) becomes significant. Maybe there exists some way of laying it out or slicing it to make it clearer...?