Independent Component Analysis (ICA) in Matlab

Independent Component Analysis (ICA) in Matlab:

While performing data analysis tasks, it is generally assumed that the feature attributes are independent of each other. But data may not be available in such format. So it is very important to convert the feature attributes independent of each other. This can be done with a simple function for ICA in Matlab. This function will reduce the data into required number of feature attributes. 

ICA technique basically removes the linearity factor among the feature attributes. Matlab does not provide in built function for ICA. Various packages [1][2][3] of ICA in Matlab are available. It can also be termed as a method for dimensionality reduction. ICA basically projects the data into N number of dimensions. Note that ICA can only separate linearly mixed vectors. ICA assumes that the vectors are non-Gaussian in nature. 

[1] http://www.mathworks.com/matlabcentral/fileexchange/38300-pca-and-ica-package

[2] http://research.ics.aalto.fi/ica/fastica/

[3] http://jim-stone.staff.shef.ac.uk/bookmatlabcode.html

Policy Iteration for Grid World Games

In my previous post, I discussed about Value Iteration algorithm for Grid World Games. Policy iteration is another algorithm which is commonly used for solving MDP problems. Here I will be discussing Policy Iteration algorithm for Grid World Games. 

Consider the same problem in the previous post. Both Value Iteration and Policy Iteration algorithms are initialized to arbitrary and random values which are far from optimal solution. These initialization do not take into account the specification of all states S and actions A which may be detrimental to the speed of convergence. For each iteration, the algorithms explore all the set A of actions, defined in the MDP, to improve the current policy, this may be causing unnecessary calculations. The transition model used is fixed, but, as we shall see later, in robotics a model of this type can degrade the results and not be in agreement with the theory of MDP.