Closed abcorrea closed 4 years ago
The idea we are currently trying is the first one. We will run experiments to check the overhead this causes as soon as the grid is back. The main reason for choosing the first one is because this is the minimal sound modification to the algorithm described in the paper.
Yannakakis' algorithm is not safe when we compute inequalities of non-distinguished variables before processing the whole join tree. We would like to test alternatives that make the algorithm safe and complete. The possibilities are:
1) Add variables in inequalities to the distinguished variables 2) Perform projection only at the root node of the join tree
In the ICAPS paper experiments, this seems to not be an issue for coverage.