/race-the-machine

Primary LanguageJava

Race the Machine

A Machine Learning project for CSCI489

Concept

The idea of the project is to develop a neural network to play "Race the Sun". An image processor will first grab and process a snapshot of the game's screen, and then the results of the processed image - which will be basic geometric shapes as well as positional and size differences from the last processed image - will be fed into a neural network.

After, the system is told how well the network performed, allowing it to perform "natural" selection based on the score as a fitness value. The score is a natrual choice as it represents both the distance traveled, and the extra points collected during that time. Because of multipliers, a well performing AI could see quadratically increasing score, further helping the system to decide.

Building our Code

To build our code, you will need Java and Gradle installed. If you are not running on Linux or Windows, you will also need to build OpenCV for your platform. The command we used was cmake -DBUILD_SHARED_LIBS=OFF -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTS=OFF .. Note: You will need ant installed to build OpenCV for Java.

Unfortunately due to updates, you may need to rebuild on linux to run depending on what system you use. All tests were run with Arch, so chances are we used newer versions of some libraries.

Once built, stick the binary library in the root of race-the-machine.

To run OpenAI Gym, you will need to install it with pip; we recommend installing both gym and 'gym[box2d]'. Take a look at their github page for more information.

Running our Code

Currently HyperNEAT does not fully work, and AHNI is not provided with reliable score information making it incapable of learning as well. That said, to give a test to our NEAT algorithm, there are two main tests worth trying. The first is the XOR test, which is under the neural-net/plu.teamtwo.rtm.experiments package.

To run any of the NEAT OpenAI Gym tests, first run the PythonServer in the same experiments directory, and then run python/openai_gym_neat.py. In the python file, you can change out what environment is being tested.

Unfortunately Py4J in its current configuration requires that only one server-client pair is run at a time, and if you stop one, make sure to stop both because the Java server will not reset its information otherwise, even if it accepts a new connection.

To run the Race The Sun portion which will demonstrate graphics processing, run the client/plu.teamtwo.rtm.client.Main class. If you get an error when trying to run it, it is probably because you need to compile OpenCV for your system.

Tuning Parameters

There are a lot of parameters that can be tweaked within the system, unfortunately we have not gotten around to creating a configuration file yet, so I will list the locations to check for the constant parameters - note that all of these are under the neural-net/plu.teamtwo.rtm package.

  • neat.GAController is the main set of configurations, and control the environment the individuals and species exist in.
  • neat.Species has configurations for determining species behavior and or membership.
  • genome.graph.GraphEncoding has configurations for the base NEAT encoding, these parameters will also alter how HyperNEAT works as a GraphEncoding is used by it internally.
  • genome.graph.MultilayerSubstrateEncoding has configurations for HyperNEAT only parameters.