Grading bouldering problems is a very important part of bouldering. It’s also a very demanding task, the success of which is determined by the grader’s experience and consistency when grading. This thesis compares different approaches to encode bouldering problem properties and the accuracy of machine learning models, such as random forest, linear regression and k nearest neighbours, when grading bouldering problems, in hope that the machine learning models prove to be sufficiently accurate to appropriately grade an ungraded bouldering problem.
Results are presented in different phases. A new set of unused or newly created features is added in each subsequent phase and the accuracy of learning is compared across different models for the selected subset of features. Finally a subset of generic features is used, such that it can be transferred to any climbing wall, which adds additional value to the work done in the thesis. The end result is a set of machine learning models, which, when grading an ungraded bouldering problem for an arbitrary climbing wall, miss the correct prediction for 1.69th of the grade.
|
