[Solved]: Dynamic Learning (Machine Learning)

Problem Detail: 

I have been given a task to dynamically learn the optimum value of a parameter in a Heuristics Filtering Algorithm used in a tool.

The accuracy of the tool increase as the value of K in the heuristics filtering algorithm increases. How ever a lower K leads the tool towards a low resource consumption (memory, time and temporary storage).

How can I dynamically learning the optimum value of K using Machine Learning techniques so that the tool can dynamically generate an optimal value of K.

I have been looking into Reinforcement Learning and learning from online resources. First, the problem has to be modeled as a Markov Decision Process (s, a, r, s'). I got stuck at what my State Space s should be. My set of actions a (they represent the value that k should be) are {10,20,30...,100}. Is Reinforcement Learning the way to go in the first place? If not, what other machine learning techniques do you suggest? I'm not into AI but I enjoy solving problems

Asked By : aig

Answered By : NietzscheanAI

First some questions:

• Can $k$ only take on 10 possible values?
• From what you've said above, the measure of the 'quality' of a given value for $k$ would be an n-tuple e.g. $(accuracy,memory, time, storage)$, right?

So 'dynamically learning a value for $k$' involves learning a good choice based on $L$, the list of n-tuples for previous values of $k$.

Here are two very simple schemes:

If your problem is genuinely multi-objective, then you can always just pick the associated $k$ at random from the Pareto-front of $L$.

If your multiple objectives can be combined into a single quality value, then fit a differentiable curve (e.g. a polynomial) to these values, then symbolically determine the optimum of that curve and the associated value of $k$. Most computer algebra packages support the functionality for this, as does the Apache Commons Math library for Java.

Best Answer from StackOverflow

Question Source : http://cs.stackexchange.com/questions/50898

3.2K people like this

 Download Related Notes/Documents