RAFF - Robust Algebraic Fitting Function

This package implements a robust method to fit a given function (described by its parameters) to input data. The method is based on the LOVO algorithm [1] and also in a suitable voting strategy in order automatically eliminate outliers.

Potential use of this package is in statistical problems which depend on fitting functions or in more specific fields like computer vision for problems related shape detection.

For a brief description of the potential of this program and the difference between related strategies in the literature, like least square, consider the following dataset

A=[-2.0 5.0; -1.5 3.25; -1.0 2.0 ; -0.5 1.25; 0.0 1.0 ; 0.5 2.55; 1.0 2.0 ; 1.5 3.25; 2.0 5.0 ;]

Let us assume that we wish to find a fitting function given by parameters x[1], x[2] and the model

f(t)=x[1]*t^2+x[2]

where f(t[i])≈A[i,2] in the least square sense. When we apply a least square strategy we obtain

and when we apply RAFF strategy we get

Note that RAFF detects (automatically) the model f(t)=1.0*t^2 +1.0 that is, RAFF detects the point (0.5,2.55) like an outlier and excludes the influence of this point for determine f(t). Consequently, RAFF detects the number of possible outliers and in the same time, which their are. This fact distinguishes RAFF from other techniques like [1] or classical least square.

Other useful examples are found in the documentation and also in directory examples/.

Installation

RAFF is a package registered in the General repository. Consequently, in order to install RAFF, just type in REPL:

julia julia> ] add RAFF

Advanced

Several files with test problems are provided inside directory test/test_problems. We also provide some scripts to create new random test problems, visualize and solve them using RAFF. Those files can be found in test/scripts. For a detailed explanation of each script file, please refer to the documentation.

Contributing

Any suggestions or contributions to RAFF are very welcome. RAFF is a open-source software available through the MIT license, so users are free to create their own versions of the software. Users are specially encouraged to participate on the project in the following ways:

• Modifications to the code can be submitter as pull requests to the branch master.
• Bugs, suggestions and questions can be raised as issues in the project's Github page.
• General doubts can also be sent directly to fncsobral at uem dot br.

Performance

The following table outlines the results found by the current version of RAFF in the test set.

| Name | Dim. | N Points | N Outl. | Found | Correct | Time (s) | Status | Solution | | ---- | ---- | -------- | ------- | ----- | ------- | -------- | ------ | -------- | | C1 | 4 | 10 | 2 | 2 | 2 | 0.0006 | 1 | [ 1.999e+00, 1.142e-02, -4.023e+00, -1.053e+01] | | C2 | 4 | 10 | 2 | 1 | 1 | 0.0005 | 1 | [ 1.575e+00, 2.635e+00, 3.434e+01, -7.821e+00] | | C3 | 4 | 10 | 2 | 2 | 1 | 0.0005 | 1 | [ 4.221e-01, -1.596e+01, 2.388e+01, 2.146e+02] | | C4 | 4 | 10 | 1 | 1 | 1 | 0.0005 | 1 | [ 2.701e+00, 2.556e+00, -5.914e+01, -1.021e+02] | | C5 | 4 | 100 | 1 | 9 | 1 | 0.0210 | 1 | [ 2.114e+00, 1.160e+00, -6.590e+00, -1.971e+01] | | CI1 | 3 | 25 | 2 | 2 | 2 | 0.0020 | 1 | [ 9.741e-01, 9.769e-01, 5.020e+00] | | E1 | 3 | 20 | 1 | 1 | 0 | 0.0263 | 1 | [ 3.271e+03, -9.884e+01, 5.507e+01] | | E2 | 3 | 10 | 1 | 2 | 0 | 0.0110 | 1 | [ 2.137e+03, 2.760e+03, 2.369e+02] | | E3 | 3 | 100 | 1 | 1 | 1 | 0.2297 | 1 | [ 9.590e+02, 9.632e+03, 3.940e-01] | | E4 | 3 | 100 | 10 | 1 | 1 | 0.2308 | 1 | [ 1.139e+03, 9.370e+03, 4.206e-01] | | LOG1 | 4 | 10 | 1 | 0 | 0 | 0.0042 | 1 | [ 1.479e+03, 2.201e+03, 6.611e+01, -2.075e+02] |

Credits

• Francisco Sobral (Leader)
• Emerson Vitor Castelani
• Ronaldo Lopes
• Wesley Shirabayashi

The authors of this package were sponsored by Fundação Araucária, project number 002/17 - 47223. If you would like to cite this package, please use

Castelani, E. V., Lopes, R., Shirabayashi, W., & Sobral, F. N. C. (2019). RAFF.jl: Robust Algebraic Fitting Function in Julia. Journal of Open Source Software, 4(39), 1385. https://doi.org/10.21105/joss.01385

The following paper describes the theory and several comparison tests

@article{Castelani2021, author = {Castelani, Emerson V. and Lopes, Ronaldo and Shirabayashi, Wesley V. I. and Sobral, Francisco N. C.}, doi = {10.1007/s10898-020-00970-4}, journal = {Journal of Global Optimization}, title = {{A robust method based on LOVO functions for solving least squares problems}}, url = {http://link.springer.com/10.1007/s10898-020-00970-4}, year = {2021} }

References

[1] Andreani, R., Martínez, J. M., Martínez, L., & Yano, F. S. (2009). Low order-value optimization and applications. Journal of Global Optimization, 43(1), 1-22.