Develop an offline system that is able to perform level design for classic computer games based on (1) representation transformation and (2) transformer usage. The prototype should solve the problem for the classic arcade game Hocus Pocus (HP), as it provides a level editor and should allow relatively simple re-engineering of the level format (and is one of the supervisor’s favorites). Practically, the outcome of the project should be: (a) a transformer between the binary HP format and, for example, some JSON format; (b) selection and, if necessary, transfer learning of an SLM for the transformation of a story prompt into a JSON level representation (or sequence of levels), (c) prompt engineering, so that the end user can create (sequences of) levels by the provision of a simple plot (e.g. “A ride through Dante’s inferno”).
Requires good knowledge in natural language programming and text understanding. If performed as a diploma thesis, must include a user-based evaluation. AI-based code generation is not allowed for this task (which is the default for such tasks).