Raw EC Notes on Ringo Spectral Problem

Intro

Finshed Haskell code. It's a naked, raw EC. No frills but I have full control of everything.

Tried "Hail Mary" on 20 PCA cols and got ..

%%bash
./go pcacolsPlusR.csv

[1 of 1] Compiling Main             ( Main.hs, Main.o )
Linking main ...
Running ..
Best guy -> 1084.844018614716
            [1.372717358007764,0.20492317564052162,-1.2202443821808244,2.5893182758257378e-2,8.024132132649084e-2,-2.247259035817572e-2,-3.836608412037176e-2,-8.765290787334143e-2,0.12746910297841527,-2.2691987748598083e-2,-4.994802038114349e-2,0.11448778257051925,-2.848859187450247e-2,3.8135548272204807e-3,5.4276008323300914e-2,-3.434923087812508e-2,9.824260783998357e-3,-5.324306464761895e-2,3.7524530449752354e-2,1.6792042616536107e-2,3.822572072354418]

Best fitness -> 1084.844018614716
-1.927838763495747   30.93878707089
-20.086027008276456   12.8845940851425
4.635401043043105   37.5778502542997
43.721313835107374   76.6438399408294
-17.585455012121677   15.3834068487984
-20.41530283825956   12.555091638161
-25.870807996592912   7.0995415695064
9.67033886365778   42.5994058435662
-23.487184813264516   9.51402656864492
16.886842660691038   49.7274114557595

Done ...

real	3m14.101s
user	3m15.052s
sys	0m10.600s

Messy printout but it says ...

• The best MSE was 1084 .. Bad
• The $\alpha's$ go between -8 to + 4 .. Bad
• It took over 3 minutes to run 1000 generations with a population of 300 .. Fair

Next Tried only 3 PCA columns

%%bash
./go only3.txt

Running ..
Best guy -> 1084.8730409295079
            [1.3692292732107871,0.2050705610860962,-1.2194775169144898,-0.11885742519850402]

Best fitness -> 1084.8730409295079
-1.8480463768243216   30.93878707089
-20.214267963838935   12.8845940851425
4.725178859665317   37.5778502542997
43.83008487090973   76.6438399408294
-17.489545585098863   15.3834068487984
-20.58318644352925   12.555091638161
-26.054799787611643   7.0995415695064
9.757073058565584   42.5994058435662
-23.51262982050006   9.51402656864492
17.23486848001705   49.7274114557595

Done ...

real	0m50.262s
user	0m51.448s
sys	0m5.796s

No better. The MSE is still 1084

Try a Tiny Test Matrix

%%bash
cat testfile.txt

 1  2  3  20
 4  5  6  47
 7  8  9  74
10 11 12 101

The $\alpha's$ for this are, $[2,3,4]$

Let's run it !

%%bash
./go testfile.txt

Running ..
Best guy -> 3.1297530799026013e-3
            [3.3374362396627433,0.1956540460027632,5.454754388981784,-0.11885742519850402]

Best fitness -> 3.1297530799026013e-3
20.093007498613623   20.0
47.05654152255549   47.0
74.02007554649737   74.0
100.98360957043924   101.0

Done ...

real	0m4.244s
user	0m4.812s
sys	0m0.940s

A bit better ...

• MSE is 0.003.
• Took ony 4.3 seconds to run.
• Appears to show that the code is working.
• $\alpha's$ are between approximately 5 and -0.1. ... Bad

So now it appears that we need to reduce the search space by limiting the $\alpha's$ somehow. I'll have to think about this.

INotes

• Got it all working on Big Iron .. It runs slower! On the good side I changed the program to input random seeds from the command line so I can run 6 different seeds at the same time (one per core) without threading magic. I wrote a script to do this automatically. It hasn't melted the CPU yet.
• Even though I started off each individual with $\alpha's$ between zero and one it moved out of that range (often negative).