Skip to content

NANO: Noise-Attenuating Neuron Online: real-time machine learning that cares about the noise in data.

License

Notifications You must be signed in to change notification settings

Bill-Armstrong/NANO

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

58 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Running the programs

First of all you have to build all the projects. I use the community version of Visual Studio 2019. Then look at the properties of both think and Q learning projects and in the Configuration properties, Debugging set the working directory to ..\..\..\Working\. ( If you are reading this as a .txt file and not an .md file, the double backslashes should be read as single ones.)

After seeing how these programs work, you can try your own data files (without header lines!), and make your own program, e.g. in Python, to run the NANO library using the NANO DLL(in the X64/lib directory). I'll try to get some documentation to help with that asap.

To try think.exe

Look at the properties of the think project in VS, particularly Debugging. That is where the data file is named and the rest of the command line is defined. The program takes input data from the directory called Working in directory NANO and produces output there (see Project Properties below to make sure the working directory is set correctly). The file ExtendedTR.txt has an additional column with the learned ouput. For the sine-cosine problem, a function which has two inputs, you insert a scatterplot (using, say, MS Excel) from the right-hand two columns that shows how closely the noisy data is fitted.

One example input is VeryNoisy.txt which has a huge level of noise applied as the signal gets smaller. The goal of the program is to try to deal with a variable level of noise in data. The output is in the file ExtendedTR.txt which you can examine in a spreadsheet. In MS Excel, for the VeryNoisy.txt input data, you open the file, highlight the rightmost two columns and insert a line chart which show the noisy data input and the NANO output ( which extends the input data by one column).

Project properties: Here is a typical command line for think.exe which contains some arguments for the program which can be placed in the Debugging > Command Arguments property: NoisySinCos20000.txt 3 20 -90 1 2 3 These are, in order, "Data_file_name nDim nMaxEpochs dblMSEorF List of columns used". nDim is the number of inputs to the function to be learned plus one (for the single output). An epoch is one pass through the data in the buffer. In the last of nMaxEpochs in a call to Train, the hyperplanes which form pieces of the learned result can split in two. MSEorF shows the level of mean square training error of a piece above which it can split. If MSEorF is negative, the program does an F-test based on the estimated level of noise to decide on splitting. The alpha probability of the F-test is the negative of the MSEorF entry. The alpha values are accurate from tables only for 25, 50 and 75. The rest are computed by geomentric interpolation. The last arguments are a list of the columns that are used in forming the ALN function, the last one being the column with the desired output. The columns are numbered from the left starting at one. For the VeryNoisySine.txt problem, column 2 is the noiseless result, so you have to skip that column, e.g. "VeryNoisySine.txt 2 20 -55 1 3".

To try Q_learning.exe

This shows NANO performing Q learning for the inverted pendulum problem. Just run the program (say inside Visual Studio "without debugging" ). It does a few iterations of Q learning then asks you for an integer number of iterations you want to do. You can give it 100, 1000, 20000 or more. At the start 1000 iterations takes on the order of a minute or two. As the program learns the Q function, the ALN tree grows and may end up with hundreds or thousands of leaf nodes. Then each iteration takes longer. Each leaf node computes a flat surface defining a small part of the function surface. After the specified number of iterations, the program asks again how many more iterations you want. If you input 0, then that finishes the program and you can open the file qmotion.txt in the directory "Working" using a spreadsheet like Excel. Then you highlight the B column and insert a line chart which shows the angle of the pendulum over time. The C column shows the angular velocity. Then comes the control given. If you want to show the time in the first column, you can use scatterplots of the A column with others. There are just three possible torque values which can be used: maximal clockwise torque (negative), maximal counter-clockwise torque and zero torque.

The qmotion_EXAMPLE_OUTPUT.txt result needed 21000 iterations, but I suggest you try 1000 additional iterations at first. Experimentation can show how to do the job with fewer iteations.

Project properties: change the working directory for this project to the same as above for think.

About

NANO: Noise-Attenuating Neuron Online: real-time machine learning that cares about the noise in data.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages