search

nazaruka / gym-http-api

NSGA2-based Sonic agent + experimental code
MIT License
1 stars 1 forks source link

Define flexible architectures #11

Open schrum2 opened 5 years ago

schrum2 commented 5 years ago

Once #8 is done, you will be able to run any algorithm in any domain, but what will be the architecture of the network that is used to learn?

Provide a way to specify the network architecture as a list of numbers, where the number of elements is the number of dense layers, and the specific values are the number of hidden neurons.

So, if you sent a parameter [10,5,3] then a network would be created with some number of inputs defined by the domain, then 10 hidden neurons, then 5 hidden neurons, then 3 hidden neurons, then a final output layer where the number of neurons depends on the domain.

schrum2 commented 5 years ago

Note: You might actually want to do this in conjunction with #8

schrum2 commented 5 years ago

Please provide an update on progress for this issue. It is still relevant

schrum2 commented 5 years ago

I would like to have the ability to define some different architectures for the Sonic PPO agent as well. This doesn't necessarily have to be smoothly integrated yet ... but there should be a clear access point in the code where the architecture can be easily changed. You need to find where the architecture is defined in the PPO code first.

schrum2 commented 5 years ago

Although this is no longer something that we want to be able to do via the command line for different domains, we would still like to have an easy way to define different architectures for the PPO Sonic agent. So, this issue is sort of being hijacked.

If you uncover a way to easily run that PPO Sonic agent with a variety of different architectures, mention it here. Otherwise, ignore this issue.