Using Combinatorial Optimization and Game Theory concepts to build a general AI game bot Using Neuroevolution and RL to train an agent, with learnt policies being represented dynamically in a memory storage.
In contemporary RL, policy search methods can be broken down into gradient-based methods, also known as policy gradient methods, and methods that do not rely on the gradient. Gradient-free methods include evolutionary algorithms. ES and GA are similar terms referring to this very approach.
This project is likely to end up being the virtual agent that the ROSCOG project aims. For neuroevolutionary approach with implicit environment representations, see: (https://github.com/blackvitriol/Genetic_Programming)
Collecting resources:
- http://deap.readthedocs.io/en/master/
- https://github.com/glample/Arnold/blob/master/README.md
- https://github.com/SerpentAI/SerpentAI
Environment: Using MuJoCo: Modeling, Simulation and Visualization of Multi-Joint Dynamics with Contact This is a physics engine to be used in conjunction with dm_control 'The DeepMind Control Suite and Control Package"
Work on :
- Desktop games (including browser games).
- OpenAI bots that are being rendered
Before running: Get a key from MuJoCu's website and place it in the bin folder, then follow: https://github.com/deepmind/dm_control
export LD_LIBRARY_PATH=~/.mujoco/mjpro150/bin sudo apt install mesa-utils ldd $(which glxinfo) | grep libGL.so LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so.1.13:/usr/lib/nvidia-384/libGL.so.
Platform: Ubuntu 16.04