Policy evaluation

[Evil-Incorporated]

In the paper you mention that the policy evaluation is done by sampling K random sample sequences. First, I am just curious as to how big of a K are you using, second with high model horizon of like H=16, how is that you are sampling a sequence that will even have anything close to optimal behavior, so essentially my question is when you say you are sampling K randomly sampled action sequences, are you sampling random actions for each time step because if you are then for large model horizons you need a huge value for K to get to a action sequence that might be close to optimal, or are you doing something else and if so please tell me what that is.