A Structural Approach to Dealing with High Dimensionality Parameter Search Spaces

Abstract

In the field of robotics, searching for effective control parameters is a challenge as controllers become more complex. As the number of parameters increases, the dimensionality of the search problem causes results to become varied because the search cannot effectively traverse the whole search space. In applications such as autonomous robotics, quick training that provides consistent and robust results is key. Hierarchical controllers are often employed to solve multi-input control problems, but multiple controllers increases the number of parameters and thus the dimensionality of the search problem. It is unknown whether hierarchies in controllers allows for effective staged parameter optimisation. Furthermore, it is unknown if a staged optimisation approach would avoid the issues high dimensional spaces cause to searches. Here we compare two hierarchical controllers, where one was trained in a staged manner based on the hierarchy and the other was trained with all parameters being optimised at once. This paper shows that the staged approach is strained less by the dimensionality of the problem. The solutions scoring in the bottom 25% of both approaches were compared, with the staged approach having significantly lower error. This demonstrates that the staged approach is capable of avoiding highly varied results by reducing the computational complexity of the search space. Computational complexity across AI has troubled engineers, resulting in increasingly intense algorithms to handle the high dimensionality. These results will hopefully prompt approaches that use of developmental or staged strategies to tackle high dimensionality spaces.