You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
… of terminal states (and does not use caching) and MSE is used for heuristics (and caches values). This makes sure that the best expressions returned by the algorithm are certain to have their parameters optimized.
ff40e81 separates the calculation of MSE (for heuristics) and the calculation of MDL (for model selection) but is not an ideal solution.
The main problem here is that two expressions that might be semantically equivalent and therefore have the same error can have different MDL values. Returning a cached MDL value for expressions with the same semanic hash would therefore be incorrect, while returning cached MSE values is fine.
Actually, the same issue concerns the de-duplication of search states (semantically equivalent states may have different MDL). However, this is not a problem as long as states are visited in order of their length (as in breadth-first and beam search).
The best solution would probably be to store the MDL of the shortest expression (assuming that a shorter expression probably has a better MDL).
No description provided.
The text was updated successfully, but these errors were encountered: