What is the goal?

[biggzlar]

https://github.com/IntelVCL/DirectFuturePrediction/blob/b4757769f167f1bd7fb1ece5fdc6d874409c68a9/DFP/future_predictor_agent_advantage.py#L86

Maybe this is just me, but I find this section as described in the paper highly confusing. What exactly is g? Is it the objective coefficients, the temporal coefficients (since it is supposed to have the same dimensionality as f, not as f_i), a combination of the two as this implementation assumes or the actual g * f?

[avijit9]

@biggzlar did you figure it out?

[biggzlar]

@avijit9 Hey there, I'm afraid not. Sorry. :/

[avijit9]

I think I have figured it out. Check the make_loss function
per_target_loss = my_ops.mse_ignore_nans(pred_relevant, targets_preprocessed, reduction_indices=0) loss = tf.reduce_sum(per_target_loss)

That means you take mean across samples (as given in equation (4) and then just take sum across the time-steps. You have temporal coefficients (0,0,0, 0.5, 0.5, 1) and goal defines your objective.