We describe an attempt to generate complete arcade games using the Video Game Description Language (VGDL) and the General Video Game Playing environment (GVG-AI). Games are generated by an evolutionary algorithm working on genotypes represented as VGDL descriptions. In order to direct evolution towards good games, we need an evaluation function that accurately estimates game quality. The evaluation function used here is based on the differential performance of several game-playing algorithms, or Relative Algorithm Performance Profiles (RAPP): it is assumed that good games allow good players to play better than bad players. For the purpose of such evaluations, we introduce two new game tree search algorithms, DeepSearch and Explorer; these perform very well on benchmark games and constitute a substantial subsidiary contribution of the paper. In the end, the attempt to generate arcade games is only partially successful, as some of the games have interesting design features but are barely playable as generated. An analysis of these shortcomings yields several suggestions to guide future attempts at arcade game generation.