The examples of A2C implementations illustrate the different capabilities of the library and act as a tutorial. To understand all the different aspects, you have to read the examples in the following order:
- Mono CPU - The simplest implementation
- Mono CPU 2 - Splitting the agent in three to illustrate the modularity of
salina, seperating the critic from the policy. - Multi CPUs - Making use of the
RemoteAgentto execute over multiple CPUs in parallel - Asynchronous Evaluation - Making use of the asynchronous functionnality to evaluate policies in parallel without slowing down the learning
- GPU - Making use of a GPU for the loss computation
- Complete - The reference implementation of A2C used for benchmarking. It contains different policies architectures to illustrate how one can define complex policies.