[tomrod]

particle swarm optimization, genetic algorithms, and tabu search/heuristic search are some items I'm aware of to force out of local optimum. Using Halton sequences can also help cover the space for search initialization, versus simple random draws in a space.

[zerodensity]

From my granted limited understanding Adam is basically gradient decent combined with heuristic search.

[yobbo]

Adam is somewhat analogous to an audio compressor on the gradient "signals".

(Edit: eh sort of ...)