1
$\begingroup$

sorry if this question is out of place. I'm a begginer to machine learning, and I have use for a technique, and I don't even know where to look. The problem is:

  • I have 5 features which are real valued (Parameters in a deterministic simulation).
  • This features determine two aspects of the instance (model solution). Its feasibility (binary) and some measure of likelihood given certain experimental data (only for the instances that achieved feasibility).

Since I want to avoid generating "infeasible" combination of features, what I devised was an algorithm that iteratively does the following:

  • Generate Nc candidate feature vectors
  • Evaluate Feasibility and Likelihood for each
  • Find linear combination of features that involves a compromise between least amount of features / holds largest cluster of feasibility. Add this as constraints to the feature vector generation.

In short, it detects and iteratively refines "simple" constraints that once added to the feature vector generation "guarantee" its feasibility to save computational time evaluating combination of parameters that lead to infeasible models. Afterwards, they could be tested by inverting them and looking for other "regions" (if any) of the feature vector where the model is feasible.

Any name of techniques and references I might look for ?

Thanks!

$\endgroup$
2
  • $\begingroup$ In the set up of your classification problem does the binary feature (the response/ label) that you are trying to predict has categories that roughly mean: "feasible" and "not-feasible"? $\endgroup$ – Nitesh Jun 1 '15 at 21:49
  • $\begingroup$ Yes, indeed, those are exactly the categories I want to "predict". I've been furthering my research (will edit OP later): I want to extract rules on the feature space ("minimum" complexity linear combination of features values is enough). Is there something like a tree of linear combinations ? $\endgroup$ – cladelpino Jun 2 '15 at 23:00
1
$\begingroup$

Your algorithm is very similar to a Genetic Algorithm, which has almost the same 3 iterative steps.

You would be missing a Termination condition to end the search for an answer, once your reach an acceptable one.

You can probably adapt your algorithm to fit under a formal GA search heuristic.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.