Time parameters used during the experiments produced in the article?

[PFery4]

I would like to know what value of the truncated_backprop_length parameter was used for producing the results displayed in the publication.

The train.sh script hints at the value being either 1 or 12. Was there a fixed value chosen for all experiments across the 5 datasets from UCY/ETH?

On this topic, I would just like to know for sure, the observation time T_O mentioned in the article is 8, however the few trainings shown as examples in the train.sh script indicate either 0 or 7. Was the value of prev_horizon chosen as 8 in your experiments?

Finally, could you tell me what is the reason behind the implementation of both truncated backpropagation through time truncated_backprop_length and the choice of a set observation time prev_horizon? I understand that they both operate in different manners in the code. But as implemented in this way currently, the input feature vectors for the past trajectories are essentially shifted versions of the velocity signal of the query agent, as such: batch_vel.shape = [batch_size, truncated_backprop_length, input_state_dim * (prev_horizon + 1)]

Let:

truncated_backprop_length = 3
prev_horizon = 4

Then:

batch_vel[0] = 
[[vx_-2, vy_-2, vx_-3, vy_-3, vx_-4, vy_-4, vx_-5, vy_-5, vx_-6, vy_-6],
 [vx_-1, vy_-1, vx_-2, vy_-2, vx_-3, vy_-3, vx_-4, vy_-4, vx_-5, vy_-5],
 [vx_0, vy_0, vx_-1, vy_-1, vx_-2, vy_-2, vx_-3, vy_-3, vx_-4, vy_-4]]

This results in an input feature vector containing duplicate values of the velocity signal across time truncations within the batch (ie, in the second dimension of the batch).

[bbrito]

For the results presented in the article I have used prev_horizon as 8. This is the standard used across all prediciton papers. For the truncated_back_propagation I have tried different parameters. If the dataset has large enough trajectories the value that works best is to set truncated_backprop_length to be the same as the prediciton horizon. If your dataset only contains small trajectories you should use a lower value for the truncated_backprop_length

[bbrito]

the inspiration to use truncated back propagation comes from here