complete_with_async_eval

A2C Asynchronous Evaluation

Starting from the multi-CPUs implementation, we explain how we can implement a non blocking evaluation of the policy made on other CPUs

• It makes use of the RemoteAgent._asynchronous_call function that is a non-blocking function (when num_processes>0).
• It uses two copies of the policy agents - one for training and one for evaluation
• Note that, in that case, we can evaluate the policy in deterministic mode