How to convert this to a continuous action space?

[Shreeyak]

Impressive work! I was wondering, what would it take to convert this to continuous? Break the time step down to the fps and make a decision every frame on what force to apply in which axis and direction? Is this doable with DQN?

[Kjell-K]

Sorry for the late answer, I had strange notification settings.

Also extending our agents to continuous action spaces is straightforward by applying the same changes which the continuous-action version of DQN applies to the original algorithm. The result is DDPG and is already implemented in keras-rl.

So you have to change the action space in AirGym environment and use keras-rl's DDPG.