https://github.com/asmeurer/soft-computing-final-project

A neural network written in Python that was my final project for my Soft Computing course.
https://github.com/asmeurer/soft-computing-final-project

Last synced: 10 days ago
JSON representation

A neural network written in Python that was my final project for my Soft Computing course.

• Host: GitHub
• URL: https://github.com/asmeurer/soft-computing-final-project
• Owner: asmeurer
• License: other
• Created: 2012-12-17T04:51:04.000Z (about 12 years ago)
• Default Branch: master
• Last Pushed: 2012-12-17T04:59:33.000Z (about 12 years ago)
• Last Synced: 2024-12-31T09:28:40.999Z (13 days ago)
• Language: Python
• Size: 293 KB
• Stars: 0
• Watchers: 4
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
Neural Nework

=============

By Aaron Meurer.

This was the final project from my CSE 464 Soft Computing class.  It is a

neural network written in Python, which I attempted to use to solve the subset

sum problem.  See the report for more details.  In short, neural networks are

not suited to solve this problem (it is a NP-Complete problem), and I ended up

learning more about how to make neural networks and get them to converge to

input data than about solving this particular problem.

Info

====

The neuralnet.py file is the neural network.  To train it, run

    ./neuralnet.py patterns

where patterns is a file of patterns.  The patterns should be of the

format

    (1, 2, 3, 4), True

    (5, 6, 7, 8), False

Where the dimensionality of the tuple can be anything (but should be the

same for all patterns).  To generate patterns for the subset sum

problem, run

    ./generate_patterns.py patterns

To test the neural network on the file patterns1, run

    ./neuralnet.py patterns -t patterns1

After training, the weights are saved in the file weights.py.  You can

use this to test the network without retraining, but running

    ./neuralnet.py patterns -w weights.py -t patterns1

And it will start training from the patterns file with the weights in

the weights file.  If the weights have already converged, it will just

skip straight to training.

To do this automatically, run

    make

And the Makefile will generate two sets of patterns, train the network

on one of them, and test on the other.

There are more options to each of these functions.  Run

    ./neuralnet.py --help

and

    ./generate_patterns.py --help

to see them all.

Extensive data on the run of the neural network is saved to a file

output of the neural network is saved to the file `output` (new output

is appended so you can gather statistics on multiple runs)

To generate statistics on the outputs, run

    ./stats.py output

The files `output_presentation` and `output_report` are the outputs that I

used for the presentation and the report, respectively.

The files `or`, `xor`, and `two` are small test patterns used to test the

network.

The script `runall.sh` will run the neural network on eight different

architectures 50 times, as I did in the report.  This takes a very long

time to finish.