Maze-Machine-Learning-Demo

A Machine Learning Demo

This demo was created in eight days and as such has rough edges and lacks unit tests. I intend to revisit this project and write unit tests as well as new features.

What I Am:

This program is a demonstration of neural networks learning to solve a randomly generated maze. The training algorithm is a genetic algorithm, it kills off the least effective and breeds the most effective. The networks are "feed forward" style neural networks, meaning they have no memory or reflective behaviour.

To Compile:

C# & .NET code Compiled with Visual Studio 2013.

To Use:

• Run the application and it will present a menu to configure the environment.
• I suggest reducing the initial cut-off time to something like 400 and raising it later.
• Click the "Go!" button. The menu will close and the maze window will open.
• At any time close the maze and the menu will reopen.
• If you change properties on the menu and click "Go!" again, it will try to not regenerate the bots, however if necessary it will.
• When you are done close the maze and menu windows to exit.
• The networks will breed when the simulation reaches the cut-off time, so having an initially low cut-off time will quickly train basic behaviour necessary to solve the maze.
• It is a good idea to later increase the cut-off time, as this will weed out over fitting problems.

Training & Learning Features:

• Changing any of these options will force the bots to be regenerated and retrain from scratch.
• The "Maze Distance To Goal" option will use a path finding algorithm to score networks during breeding.
• The "Path Distance To Goal" option will allow the networks to see the fastest path to the goal.
• The "Direct Distance To Goal" option will provide the networks a value that indicates how far from the goal they are.
• Finally, "Aware of Velocity" tells the network its current speed and angle.

Planning

Refer to [Review & Roadmap](Review & Roadmap.md).