Tabu Search is a commonly used meta-heuristic used for optimizing model parameters. A meta-heuristic is a general strategy that is used to guide and control actual heuristics. Tabu Search is often regarded as integrating memory structures into local search strategies. As local search has a lot of limitations, Tabu Search is designed to combat a lot of those issues.
The basic idea of Tabu Search is to penalize moves that take the solution into previously visited search spaces (also known as tabu). Tabu Search, however, does deterministically accept non-improving solutions in order to prevent getting stuck in local minimums.
Short Term memory is based off of recency of occurrence and is used to prevent the search algorithm from revisiting previously visited solutions and also can be used to return to good components in order to localize and intensify a search. This is accomplished by the Tabu List and is also known as intensification.
Long Term memory is based off of frequency of occurrence and is used to diversity the search and explore unvisited areas of the search space by avoiding explored areas. This is accomplished by frequency memory and is also known as diversification.
The Tabu List is the cornerstone of utilizing short-term memory. This list stores a fixed amount of recently made moves. In some implementations, complete solutions are used instead of the moves used but this is not ideal if complete solutions are very large due to space limitations. Some examples of these moves are:
#machine-learning #statistics #optimization #ai #data-science