Pretty much, the way I plan to develop this is as follows:

• Develop the basic Rules/Fact classes and the like. This will start by using simple code blocks as the criteria units, but eventually will use things like Array::PatternMatcher that princepawn posted recently. (I've been talking to him on more details). The rules system would currently be explicitly defined in perl and would require perl calls to start it going.
• Develop a way to test the rules system by simply iterating over all possible rules/facts until a rule is fired.
• Develop a way to cache facts vs rules such that the above step is more efficient
• Build up a standalone text format that can be used to simply have a perl-based rules system but without using any perl code to initiate it.
• Modify the system from there.

My current thoughts is that the key data item, facts, are only arrays; they may be arrays of mixed data, or just scalars, but they are arrays. Of course, from the practical understanding of a rules-system, you shouldn't be using very complex data structures in any case, as you can always alter them to be replaced by multiple facts, but I want to leave that open in case someone wants to take the concept to advanced levels. The only time that data is introduced to the system would be by facts; while perl input can be used to generate data, the data would go out of scope at the end of the rule unless specifically put into a fact and stored away. Standard perl output can be used for any other generation, while backend stuff can be directed to another file (ie which rules have fired, with which facts, etc).

The trickiest part, for me, is making the rule triggering more efficient. Rules will have two values associated with them: priority (higher priority rules will trigger first) and probability (rules of the same priority and that all can be triggered at the current step; probability will select a random rule from this to trigger). These could change as the rules are fired, though this can lead to bad rule-based systems. In addition, some criteria for rules are dynamic, for example in my pseudo code...:

rule find-max
1:( value ?id1 ?X )
2!( value ?id2:{ ?id2 != ?id1 } ?Y:{ ?Y > ?X } )
=>
etc...
[download]

That is, the first criteria takes a rule that matches the format of "value <id> <number>". The second is a negative criteria; this is, we don't want any rule where a different ID has a larger number. Now, to check this, one must take all N facts, and do N^2 checks with the criteria (though it can be assumed that this might be closer to 0.5 * N^2, since once a fact denies the second criteria, the first rule obvious cannot match. In small cases (few rules & facts) this isn't bad, but it won't scale well to larger systems. I'd like to have a side strucutre that notes when facts match the criteria of rules in order to improve this efficiency; this is easily done for the first criteria of the example above but the second step may still require O(N^2) checks. In general, for N rules with an average of K criteria per rule, and M facts, the system scales as N*(M^K), and any possible reduction of this would help.

At this point, I've mostly got ideas, and it's a matter of implementing things in a perl-ish way (eg transferring variables from matched rules to the 'body' of the rule).

---

Dr. Michael K. Neylon - mneylon-pm@masemware.com || "You've left the lens cap of your mind on again, Pinky" - The Brain