Frog Jump AI

This is an experimental AI project of Reinforcement Learning applied to a platformer styled game.

In Frog Jump play either as a human or let a bot learn how to win a level.

You can customize the settings of the simulation and neural network to guide your bot learning process. (RL w NN + GA)

How it works?

In AI mode, a simulation with N bots will start each one with an internal neural network.

Each network can have up to 5 sensors, 4 hidden layers and 1 output neuron.

Each iteration of the simulation will run a reinforcement learning process optimized by genetic algorithms. At the end of each iteration the best bot will be mutated several times to build a new population for the next iteration. This process keeps going until the level is finished.

Adjustable params

• Number of sensors.
• Number of hidden layers.
• Size of hidden layers.
• Probability of mutation.
• Number of bots per iteration.
• Bot's horizontal detection range.
• Bot's vertical detection range.
  

Fixed params

• Hidden layers use RELU activation function.
• Output layer uses Sigmoid function.
• Decision threshold is 0.5 for output neuron.
• Mutation is element-wise with a normal range of (-2, 2).
• Initial population is randomized.

Controls

• Click/Space to jump.
• Escape to exit.

References

You can see the source code publicly and freely available on Github. https://github.com/Carlitos5336/ai-tutorials

NOTE: Recommended to play desktop version for better experience.