|title:||A fair grade: assessing difficulty of climbing routes through machine learning|
|keywords:||Machine learning, Classification, Markov Chain, Domain-Specific Language|
In the sport of climbing and bouldering, difficulty grades are important for skill assessment, enjoyment and safety. Although there are standardized scales to minimize inconsistency, the difficulty assessment is performed by humans and sensitive to subjectivity. Therefore, we have developed a tool that predicts the difficulty grades of climbing routes using machine-learning. The routes are described in a semi-natural Domain-Specific Language, which can be parsed into symbol sequences. Here, a symbol represents a climbing move. The symbol sequences are then used as inputs to a variable-order Markov models (VOMMs) based classifier. With the VOMM prediction algorithm Decomposed Context Tree Weighting (DE-CTW), we trained one VOMM on Easy climbing routes and one on Hard climbing routes. By calculating a test route’s likelihood for both VOMMs, the average log-loss, we predict if a route is Easy or Hard. We have implemented six predictor variations to vary with interpretation detail in the symbolization process. After using 50-fold cross validation on 146 climbing routes, our best performing variation performed roughly as well as a trivial classifier. Still we believe this research’s foundations are of interest for future research. We conclude with detailed explanations and proposed improvements.