Adds explicit metadata for bias intervention

[micahwoodard]

Adds explicit metadata for bias intervention to make it easier to determine autowater vs bias reward

[bruno-f-cruz]

I don't quite understand the meaning of these variables, so I will leave it up to you and the team to decide if this makes sense to implement and maintain or not.

From a technical implementation point of view, this looks great and follows the pattern we engineered. The only small note that I would have is: is there a world where you store the value of the manipulation as Optional[T], instead of a boolean flag? I am afraid that you will also need the value itself, in which case you may as well bite the bullet now.

[micahwoodard]

@alexpiet do these variable and types make sense to you? I added them only for data analysis purposes so I can implement whatever makes the most sense and is easiest for you to use

[alexpiet]

Yes, looks good. Thanks for adding it.

As an aside, I don't like the walrus operator (:=), but I'll leave that to you

[alexpiet]

From a technical implementation point of view, this looks great and follows the pattern we engineered. The only small note that I would have is: is there a world where you store the value of the manipulation as Optional[T], instead of a boolean flag? I am afraid that you will also need the value itself, in which case you may as well bite the bullet now.

I can't think of a reason you would want the bias intervention as optional. You are either doing an intervention, or you are not.

[bruno-f-cruz]

From a technical implementation point of view, this looks great and follows the pattern we engineered. The only small note that I would have is: is there a world where you store the value of the manipulation as Optional[T], instead of a boolean flag? I am afraid that you will also need the value itself, in which case you may as well bite the bullet now.

I can't think of a reason you would want the bias intervention as optional. You are either doing an intervention, or you are not.

I may be misunderstanding on kind of information you would want here. The manipulation has a corresponding value to it (T) (e.g., the distance you moved the spout, the probability you changed the reward by, etc...). A boolean will tell you that an intervention was made, but not the value of the intervention. If that is sufficient, then keep it. I am just trying to avoid a future scenario where this field will become two fields: has_value: bool and value: T.

[alexpiet]

Oh sorry, I misunderstood your question. Yes, we will want access to the value of the manipulation for the moving the stage (direction, distance). There isn't an associated value with the water delivery.

I dont know any details of how the trial generator versus how those bias intervention values are being saved, so I can't comment on how to structure these classes. I just will want access to the values and know which trial they occurred on in the trials table

[micahwoodard]

@alexpiet You'll have access to how much the stage moved from the trial.lickspout_offset_delta. If we anticipate other interventions that will move the stage beside bias interventions, it could be worth it to expose distance moved in metadata. Seems a little superfluous to me though. The main goal of these adds was to differentiate autowater vs bias interventions rewards

Trial(
         p_reward_left=1 if (self.is_left_baited or is_auto_reward_right is False) else self.block.p_left_reward,
         p_reward_right=1 if (self.is_right_baited or is_auto_reward_right) else self.block.p_right_reward,
         reward_consumption_duration=self.spec.reward_consumption_duration,
         response_deadline_duration=self.spec.response_duration,
         quiescence_period_duration=quiescent,
         inter_trial_interval_duration=iti,
         lickspout_offset_delta=lickspout_offset_delta,
         is_auto_reward_right=is_auto_reward_right,
         metadata=Metadata(
             p_reward_left=self.block.p_left_reward,
             p_reward_right=self.block.p_right_reward,
              extra=BlockBasedTrialMetadata(
                    is_autowater=is_autowater and not is_bias_intervention,
                    is_bias_water_intervention=is_bias_intervention and is_auto_reward_right is not None,
                    is_bias_stage_intervention=is_bias_intervention and lickspout_offset_delta != 0,)

[alexpiet]

That works for me