Finding and maintaining the right level of challenge with respect to the individual abilities of players has long been in the focus of game user research (GUR) and game development (GD). The right difficulty balance is usually considered a prerequisite for motivation and a good player experience. Dynamic difficulty adjustment (DDA) aims to tailor difficulty balance to individual players, but most deployments are limited to heuristically adjusting a small number of high-level difficulty parameters and require manual tuning over iterative development steps. Informing both GUR and GD, we compare an approach based on deep player behavior models which are trained automatically to match a given player and can encode complex behaviors to more traditional strategies for determining non-player character actions. Our findings indicate that deep learning has great potential in DDA.