How can one manage thousands of IF…THEN…ELSE rules?

Unfortunately, once your dimensions get large it becomes difficult for humans to visualize the optimizations, which is where Espresso[2][3] comes in handy.

[1] http://en.wikipedia.org/wiki/Karnaugh_map

[2] http://embedded.eecs.berkeley.edu/pubs/downloads/espresso/in...

[3] ftp://ftp.cs.man.ac.uk/pub/amulet/balsa/other-software/espresso-ab-1.0.tar.gz

Expert systems were particularly big in AI in the '80s, and considerably less hot now, but still widely used in industry. There are a lot of issues that come up, most of which aren't really purely technical, such as: 1) how you extract knowledge from people who might know it; and 2) how you validate that the knowledge you've extracted is actually what they know, for example by validating that it produces the same decisions that they would make; and 3) how you allow updates/revisions to the knowledge base over time. Once you have the rules and some reason to believe that they're any good, actually managing and applying them can be done through one of several rule engines, such as Drools or Jess.

The keyword "knowledge engineering" may also turn up relevant info.

For example a video game consists of millions of possibilities/results in movement based on environment - it's not approached as a collection of thousands of if/then/else statements.

In fact the movement of cows seems very much like video game coding, it needs a "cow engine".

Genetic Programming

ADTree

Random Forest

Decision Tree

Most of the work would be in figuring out the best way to gather all the data.

*opinion.

A Prolog program is essentially a list of rules that are written as an implication. Something like:

  moves_to(Cow, Location) :-
    hungry(Cow),
    current_location(Cow, OldLoc),
    food_in(OldLoc, OldFood), food_in(Location, Food),
    Food > OldFood.

I should probably write a proper answer.

edit: Added a proper answer to the stackoverflow with an almost complete introductory Prolog example. It's waiting for your upvotes :)

You start by writing a logic flowchart that you understand, then you reduce the logic flowchart to code, usually automatically. This is how a logic flowchart looks:

http://en.wikipedia.org/wiki/File:LampFlowchart.svg

Here is the article it appears in:

http://en.wikipedia.org/wiki/Flowchart

The single most important parts of this process are:

1. That you understand the logic diagram and see that it meets the program's requirements.

2. That there is a way to turn the logic flowchart into code, ideally without human intervention.

3. That the rebuild time be short between adding a new logical step and testing the resulting program.

What do the states of the cows and the landscape look like? What is changed by the if-then-else results. Does he already have his rules? Is this for a biology project or a video game, etc.

You could use a finite automaton system, an array of cow-states, a cow DSL or something else depending on the answers to these and other questions.

[insert joke about "till the cows come home"]

"How can one manage thousands of IF…THEN…ELSE rules? By developing a drinking problem"

http://en.wikipedia.org/wiki/Decision_tree

It's basically thousands of if-this-then-that statements, but dressed up into something you can reason about and most importantly train and build automagically.

I'd consider taking a look at recurrent neural networks if you're looking at your data as a time series or if its more of a static problem that doesn't take into account the last N things that came before, you might consider tree based methods or even DBNs if you can get a lot of data together (say 50000-100000 samples)

Potential pitfall: if you're using a NN based approach, you will judge results based on test performance, but you won't get as much insight into the "rules" your network has learned.

I have modeled a logistics lookup system (assign "best" courier for an ecommerce setup) using a search solution. The rules were modeled as documents in a search index and looked up using weights. Fundamentally, what I did was flatten an if-then-else hierarchy and assign weights based on (roughly) level of nesting.

The con of this model is obvious - it is at best a heuristic. However, with clever overrides built in (as a separate search index?), you can get pretty close to the ideal solution.

The pro of this model is scalability - when your rules are in millions, this system scales beautifully.

Never mind the fact that you can throw in black swan type of events into the mix or just unknown unknowns you could never think of (maybe a cow has a frog phobia and panics when it sees the frog and starts a stampede or whatever it is that cattle do).

The more modern Rete based software projects are probably also easy to work with.

Even something as simple as a Decision Tree might be enough?