Meaning of nb_prev_actions=0?

[michaelpulis-UOM]

I am using the VAEP and Atomic-VAEP Implementations. To confirm that I am understanding the implementation correctly, I set the nb_prev_actions parameter to be 0, expecting the system to fail. However, it still ran and trained correctly. So I would like to ask for clarification.

When I set the nb_prev_actions to n, and tell the model to rate the value of the actions in a match, does the model make use of n total actions to make the prediction, or n+1 actions? If it uses n actions to make the prediction, then what does it mean to have nb_prev_actions = 0?

Thanks

[probberechts]

If you set the nb_prev_actions parameter to n, and tell the model to rate the value of the actions in a match, the model makes use of n total actions to make the prediction. The idea is that each action changes the game state. For example, if you have a sequence $PASS \rightarrow s_1 \rightarrow DRIBBLE \rightarrow s_2 \rightarrow TACKLE \rightarrow s_3$ and nb_prev_actions = 2 then the features for game state $s_3$ are based on the attributes of the tackle and dribble.

If you look at the implementation you can see that values for nb_prev_actions < 1 are handled identically to nb_prev_actions = 1. Hence, if you set the nb_prev_actions parameter to be 0, you actually compute the features based on one preceding action.

I agree that it would be better to raise an exception though. I'll adapt that.

[michaelpulis-UOM]

That makes sense!

The way you explained it is the way I thought it worked, but when it didn't fail when I set nb_prev_actions to 0 confused me. I agree that it should raise an error.

Thanks