AI problems which can be solved using a state-based model are a significant portion, covering most NP problems. Although the strategies required to solve such problems are well documented and researched, the model itself is seldom analysed, being considered either secondary or too abstract to tackle in a systematic analysis. Early languages such as PDDL and the many later variants, allow users to describe a state-based model and use a type of exhaustive strategy (usually BFS) to check if that model can solve a defined problem. More recently, SMT-bounded model checking solutions have been employed, a new, generalized approach is described in this paper. Another new contribution described in this paper uses a state classifier to group generated states and thus simplify an exhaustive search for a solution. This allows us to build a complete problem-space having each state within labelled as belonging or not to any path to the goal of the problem.

Describing an unambiguous model for an Artificial Intelligence (AI) problem has been a significant topic in AI for almost 70 years, with the main goal of formalizing a natural language description to allow the computer to solve it. Nowadays, an AI problem is usually modelled as a transitional system, by following four steps: identify a representation for a problem state, describe the initial and final states in that representation, describe valid transitions together with a search strategy that looks for a path between an initial and a final state using the available transitions. This paper proposes a new language for describing AI models with the goal to generate executable code. The proposed language is capable of representing all common types of problems, and models implemented can be adapted easily to any search strategy with specific requirements (such as score functions). The language is fully described in this paper together with several non-trivial examples, while the code generation feature is work-in-progress.