BOPTESST-Gym-service

[javiarrobas]

This issue is to accommodate BOPTEST-Gym to work with the service version of BOPTEST. For the moment, this new BOPTEST-Gym-service version will be developed in parallel with master, similarly to what happens with BOPTEST-master and BOPTEST-Service. It is also expected that BOPTEST-Gym-Service will enable easier parallelization during learning which may be a very interesting feature since the emulators can run on the cloud as vectorized environments.

[javiarrobas]

Some benchmarking:

• Testing with boptest-service locally:
  

• Testing with boptest-service in a remote server (AWS EC2 t2.xlarge 16 Gb instance):
  

This means that, based on some of the tests with a co-simulation, we see that using service remotely is only ~55sec/32sec=1.7 or ~26sec/18sec=1.4 times slower. Note that using service remotely unleashes the possibility to launch several environments on the cloud with vectorized environments which may considerably speed up the time required to train RL agents (just having two environments would already be an advantage). This shows that vectorization is a very interesting feature to be implemented, especially when combining it with boptest-gym-service.

FYI @dhblum @kbenne @icupeiro.

[dhblum]

This sounds like good news indeed!