Conversation
…tion sets. This clarifies how quantities related to mixed behaviours profiles are defined for information sets which are not reached. * Beliefs, action values, infoset values, and node values are not well-defined in this case. Therefore, these functions now return a `std::optional` in C++, with `std::nullopt` when not defined. The Python equivalents now return `None` * Agent regret concepts (action regret, infoset regret, max agent regret) are well-defined despite information sets not being reachable: from the definition of the (multi)agent form of a game, these are zero. * Along similar lines to the above, in the agent Lyapunov value, the contribution to the value of unreached information sets is set to zero. Closes #446.
|
@rahulsavani Please have a look that this seems to be correct as far as conventions to you, and the docstrings have been updated sufficiently well. We might consider opening a separate discussion item that some user guide/tutorial materials on unreached information sets could be in order (also interacts with our explanation of what sequence-form-based methods return). |
| ValueError | ||
| If `action` resolves to an action that belongs to the chance player | ||
|
|
||
| See also |
There was a problem hiding this comment.
I know this is a bit OCD, but we mostly have "See Also" and a small number of cases of "See also" (lower case "a"). I'd prefer to be consistent.
It wasn't clear to me why we point to infoset_prob here under action_value but not under infoset_value, perhaps just an omission?
There was a problem hiding this comment.
Entirely because I did these on different days and wasn't consistent with myself.
rahulsavani
left a comment
rahulsavani
left a comment
There was a problem hiding this comment.
Please have a look that this seems to be correct as far as conventions to you, and the docstrings have been updated sufficiently well.
This looks good to me. Just a couple of minor inline comments on "See Also".
We might consider opening a separate discussion item that some user guide/tutorial materials on unreached information sets could be in order (also interacts with our explanation of what sequence-form-based methods return).
Agreed.
2 participants
This clarifies how quantities related to mixed behaviours profiles are defined for information sets which are not reached.
Beliefs, action values, infoset values, and node values are not well-defined in this case. Therefore, these functions now return a
std::optionalin C++, withstd::nulloptwhen not defined. The Python equivalents now returnNoneAgent regret concepts (action regret, infoset regret, max agent regret) are well-defined despite information sets not being reachable: from the definition of the (multi)agent form of a game, these are zero.
Along similar lines to the above, in the agent Lyapunov value, the contribution to the value of unreached information sets is set to zero.
Closes #446.