Algorithm selection on a meta level

Alexander Tornede, Lukas Gehring, Tanja Tornede, Marcel Wever, Eyke Hüllermeier

Link To Paper: https://link.springer.com/article/10.1007/s10994-022-06161-4

Journal Of Paper: Machine Learning (Springer)

Confirmed Open Access: Yes

Topics From Call For Papers: AutoAI (incl. Algorithm Configuration and Selection) - algorithm selection in particular

Broader Impact Statement On Ethical And Societal Implications:

# Limitations
The conclusions drawn from our work are mostly based on the runtime scenarios of the ASlib benchmark and thus are mostly limited to settings where the fastest algorithm among a set of algorithms should be chosen. Moreover, although ASlib covers a wide range of algorithmic problems, it is unclear whether our proposed approaches work well on other domains such as e.g. query optimization for SQL queries.

We discuss the limitations of the proposed approaches themselves within the corresponding subsections of Section 5.

# Ethical and Societal Implications
The impact of our proposed framework and the approaches realized within fully depends on how they are used.

On the one hand, our proposed algorithm selection approaches can be used in areas that benefit society, e.g. choosing the right algorithm for solving a traffic problem. In such a case they offer the opportunity not only to benefit society in the sense that the underlying problem is solved. Moreover, they benefit society in the sense that the problem can either most likely be solved in a better or faster way in comparison to using only a single algorithm. If solution time is the optimization criterion, employing (our) algorithm selection techniques often automatically reduces CO2 output as less algorithm runtime is required and thus automatically offers an environmental benefit.

On the contrary, the proposed methods can also be used in problematic applications, e.g. for choosing the right target algorithm of a weapon system in a specific context or any other application where one has to choose from a set of algorithms.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2107.09414/code)