I think it has to do with how bidirectional search is implemented.
The pseudocode for BFS goes something like this:
until frontier.empty?
node <- frontier.pop()
add node to explored
for each child not in expored or frontier:
if child is goal then
return path
add child to frontier
Imagine implementing bidirectional search like this:
until frontierForward.emtpy? or frontierBackward.empty?
node <- frontierForward.pop()
add node to exploredForward
for each child not in exporedForward or frontierForward:
if child in frontierBackward then
return pathForward + pathBackward
add child to frontierForward
Do something similar but now searching backwards
Basically, alternating steps of "forward" BFS and "backwards" BFS. Now imagine a graph like this:
B-C-D
/
A E
/
F - G
The first forward and backward runs of BFS would give us a state like this:
1) expandedForward = A ; frontierForward = B, F
2) expandedBackward = E ; frontierBackward = D, G
Now the algorithm picks the next node to expand from the forward frontier and it picks B.
3) expandedForward = A, B ; frontierForward = F, C
Now we run a backwards BFS and expand node D. D's child, C, is in the forward frontier, so we return the path A - B - C - D - E.
I think this is what Russel and Norvig where referring to. If the implementation doesn't consider this scenario it could give you a solution that's not optimal.
It should finish expanding all the nodes in the frontier that have the same "depth" before deciding it has found an optimal solution. Or maybe alternate the forward and backward BFS searches by layer and not by node (expand forward all nodes in layer 0, expand backward all nodes in layer 0, expand forward all nodes in layer 1, etc.)
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…