Truncation for exploration?

[AaronLiu1997]

I'm reading the paper and code, and can't follow the truncation process. Table 2 sets exploration stddev. clip equal to 0.3, so I assume that the exploration noise is clipped. However, the action seems to be selected by action = dist.sample(clip=None) which does no clipping. Instead clipping is seemingly applied during training with dist.sample(clip=self.stddev_clip). Am I misunderstanding something here? Thanks!!

[denisyarats]

This is correct, the exploration noise is only being clipped when the sampled action is being used to query the Q function.

[AaronLiu1997]

Got it, thanks! I guess I was just confused about the terminology then!