Interesting way to think about it. I found it funny because I've been working on ray tracing algorithms lately and it reminds me of that situation: They're both some form of optimization problems where you need to connect knowledge/light sources to problems/observers. Tracing from the knowledge/light source end is not efficient because you have no idea what will turn out to be important in the end, but tracing from the problem/observer end is also not efficient because in the end you have to connect your problem/observer to some knowledge/light source otherwise the entire path turned out to be futile.

Some form of bidirectional or multi-pass algorithm is likely to be more efficient here... ;-)