# Description

`robocode-svm` is a script to demonstrate Support Vector Machine (SVM) classification on outcomes of the battles of Robocode. You can check the original assignment that led to this work [here](extras/assignment.pdf).

# Installation

The script runs under Linux and requires `robocode`, `svm-scale`, `svm-train` and `svm-predict` executables located in the `PATH`. On Arch Linux (or it's derivative, like i.e. Manjaro), this can be achieved by installing [libsvm](https://aur.archlinux.org/packages/libsvm/) and [robocode](https://aur.archlinux.org/packages/robocode/) from AUR (if enabled) by following command (or equivalent):

```
pamac install robocode libsvm
```

# Usage

```
Generate the data required for SVM classification and show the accuracy
of the generated SVM model. If coordinates are provided, one visual battle
is shown to support the prediction.

USAGE: robocode-svm ACTION
USAGE: robocode-svm --battle x y [alpha]

     --battle		Shows visual run of the specified battle
  -g,--generate		Just generate the battle data
  -a,--accuracy		Show the accuracy of the SVM model,
  -c,--clean		Delete all the battle data!
  -h,--help		Show this help
```

# Details

The robots on test are **sample.Corners** (starting in center) and **sample.TrackFire** (roaming). The _gun cooling rate_ is set to **0.07**. These were obtained by experimenting and provide great outcomes on the battlefied, that can be classified easily.

## SVM kernel parameters

The SVM type used is **C-SVC** (multi-class classification), with the **radial basis** kernel. The _cost_ parameter C is **10** and the _sigma_ parameter is **2**. You can find more details about these at [libsvm home](https://github.com/cjlin1/libsvm).

## 2D visualisation

When the third parameter (initial gun orientation) of the roaming robot is omitted, we can generate _sample_ visualisation of the result of this classification. Our experiments were indicating, that this parameter has low (about 2%) significance on the outcome of the battle.

![2D visualisation](extras/visualize2D.png)

The visualisation color shows, which robot is likely to win, when the roaming one starts on given position on the battlefield. You can estimate the coordinates on the image and run script with the `--battle` parameter to prove the classification.