Hands-On Neuroevolution with Python

RL evaluation on GPU cores

In our experiment, we'll implement an RL evaluation using the TensorFlow framework on GPU devices. This means that all the calculations related to the propagation of input signals through the controller ANN are performed on the GPU. This allows us to effectively calculate more than 4 million training parameters – the connection weights between control ANN nodes – for every single time step of the game. Furthermore, we can concurrently simulate multiple runs of the game in parallel, each controlled by a different controller ANN.

The concurrent evaluation of multiple game controller ANNs is implemented by two Python classes: RLEvalutionWorker and ConcurrentWorkers. Next, we'll discuss each class.

For complete implementation ...