Instead of dslr, need to use webcam to confirm solidarity of results.

Can take 50% from the front of source, and rest from the target (after training for a few generations), to get a better population.
Hopefully getting better results.

Will help in stabilizing the results, possible converging them.

While crossover two different matrices are being used of both parents.
One is having weights between input and hidden layer and other between hidden layer and output.
So after crossover number of hidden nodes may conflict.
Possible solutions:

1. Taking min of hidden nodes and reducing both matrices to that value.
   2.Taking max of hidden nodes and adding random weights.

1. Increasing the population size to bigger values like 1000 causes the code to achieve minimum values in lesser iterations but code runs slow and could take days to reach optimum value.

2. I am deciding fitness functions using guesses...need a way to find a good function which can work well for values as small as 10^-5 or less too.

3. After some iterations minimum value nearly comes to a saturation point...this is happening because all values in population array are same.Increasing the mutation rate too much can change values in this array but that will fluctuate the minima too much.

If we remove back-propagation, (as I have removed in branch 02_wop) , GA's individual results suck!!
print results are generally like [G 10] score=(0.1161, 0.1203) [these are costs] for 1 hidden nodes
and It doesn't change with each generation, neither with each program execution
Task is to manipulate some parameters and see where the problem is.

in 04, log.txt should mention the inferences of exercise performed with commit or pull request id.

As we had originally intended, and since now Kala sir requires us to submit our codes, let's make this repository public.

Things to do:

• Copy the readme to post-midsem branch, get rid of master branch.
• Rename post-midsem as master.
• Cleanup a little bit.
• Add documentation. (code is self documented, but still)
• Integrate Travis CI, get build report functioning.
• profit???

All the types of mutation can take place simultaneously on a single chromosome, making the chromosome way out of league.

Is the modification of weights in mutation good enough?

Might help in increasing the pareto-front size.

it's just a simple nd.array for now. Change this to a dictionary (map). format is mentioned below.

` chromosome={
		                             log_reg_w : np.array([
											[blah,blah]
											,..,
											[blah,blah]
											]),
   							log_reg_b : np.array([
											[blah,blah]
											,..,
											[blah,blah]
											]),
							hid_lay_w : np.array([
											[blah,blah]
											,..,
											[blah,blah]
											]),
							hid_lay_b : np.array([
											[blah,blah]
											,..,
											[blah,blah]
											])

							
				
	} `

• make commits if you think the change made is important or lengthy, and commit name must be self-descriptive.
• This will start a long chain of 'changes', e.g. mutation, crossover, ...everywhere. Though the changes are not algorthmic, but sure is very wide-spread.
• Three pair of eyes (and hands) are desired to ensure changes are appropriate.
• Any change other than mentioned(and their neighbours ) is discouraged.
• All * abstract * issues discovered must not be addressed but first made an issue here in github.[actual issue: is the issue like print error, dimension mis-match etc; abstract issue: is the issue like 'accuracy not improving' etc]
• Issue with this issue is also to be addressed here.
• Compare results so that you know there is not a hideous error causing uncertainity . Test must be rigorous.
• Branch aside from 05_postmidsem
  P.S. Next step will be connection matrix (abstract concept). Will be seen after all is working fine with these issue fixed, do not merge connection matrix with this issue, straight away. Stepwise.