Definition
Local Beam Search
The idea of local beam search is to keep the top states (“beams”) instead of one. This is not the same as running searches in parallel. Searches that find good states recruit other searches to join them.
Quite often, all states end up on same local extreme. To mitigate this problem, it chooses successors randomly, biased towards good ones (analogy to natural selection).
States might also get too close. To address this, use a stochastic variant (akin to stochastic hill-climbing).