test.md

Overview comments on #1247 on certain things not in the java docs:

• we introduce a way to store the original first and last edge so that we do not need to store the turn cost multiple times (e.g. for every new shortcut)
• The EdgeBasedNodeContractor plays the key part of this algorithm:
  • Assume source and target nodes s and t connected like *->s-->x-->t->* and you want to contract x (center node). In order to determine if a shortcut s->t is required the contractor implements the AggressiveStrategy class that searches from all incoming edges of s to all outgoing edges of t
  • The unidirectional, edge-based Dijkstra search is done in the separate WitnessPathSearcher class, handling lots of special cases (see EdgeBasedNodeContractorTest for more details)
  • General CH concept: if there is a witness path with weight w1 and a bridge path with w2, and it is w1<w2; only then you do not need to add a shortcut when contracting x
• The current default for query time is the A* algorithm unlike the node-based CH where this is slightly slower. AStarBidirectionEdgeCHNoSOD has special methods like bwdSearchCanBeStopped
• Why is s->a->x->t a witness path and not a bridge path? Simply because we'll handle adding shortcut when s=a. This could be used for node-based preparation too, to further reduce adding shortcuts. Currently we explicitly skip x there, which might lead to a new shortcut in the case of s->a->x->t.