Neural Network Minesweeper

An implementation of the the tutorial outlined here: (http://www.ai-junkie.com/ann/evolved/nnt1.html). It is a Feedforward Neural Network with supervised learning (a simulation environment) used for the purpose of collecting mines

Dependencies

• SBT
• JDK 8+
• Scala 2.11

Tests

• All: sbt test
• Individual: sbt 'test-only org.danielholmes.smartsweepers.nn.NeuralNetSpec'
• Individual continuous: sbt ~'test-only org.danielholmes.smartsweepers.nn.NeuralNetSpec'

Running

sbt run

Observations

I've included a few different fitness strategies. They require some fine tuning to the environment - especially the avoidance fitnesses. For example a fitness to collect mines and avoid rocks - if you punish rock hitting too much then a spin on the spot strategy is developed (same as the simple avoid only fitnesses)

TODO

• Move params.ini and fitness selection into a UI
• Graph negative fitnesses better
• Try a mutable Simulation vs immutable/fully functional to see performance differences