Spaces with both Continuous and Discrete Outputs


#1

I would like to create an agent with an action space consisting of both discrete and continuous outputs. (i.e. similar to how a game controller has discrete buttons and a continuous joystick). The cleanest way I can think of implementing this is by creating a gym environment that uses either a dict_space or a tuple_space for its action_space.
However, if I do this, things get messy when creating the probability distribution tensors in my agent for a mixed space. Mixed spaces (dict_space and/or tuple_space) don’t seem to be supported by baselines/common/distributions.
Has anyone found a neat way of having a mixed action space? (using OpenAI’s PPO2 and gym) I hope this question is relevant/useful!