Dmitry Tyshlangov

    The simulation of agents in three dimensional environment with continuous state and action spaces. Each bot is driven by feed-forward neural network. The learning technique is genetic algorithm. 

    All the bots perform simultaneously, this approach is called coevolution (Marco Wiering Reinforcement learning State-of-the-Art, p. 339). In this simulation agents try to find an optimal behavior to get as much food as possible.

​

    You can easily experiment with model, for example adding moving food, adding more actions to the agent, constraining the “vision” of agents (ex. add Abs() to the angle input, that means that agent can't distinguish left and right). Look Help for more detailed info.

​

    Features:
        - commented source code
        - classes for feed-forward neural network and genetic algorithm
        - the complex reinforcement learning task environment

​