A prototypical fuzzy logic programming solver that performs inference on (ground & tight) input answer set programs using logical neural networks (LNNs).
In more detail, an answer set program is first being translated into its corresponding Clark's completion, whose abstract syntax tree provides the LNN that will be used to perform inference. Inference is performed within the LNN by
- first forwarding the truth bounds of atoms (leafs) using an upward pass;
- in order to tighten the truth bounds of atoms according to the assumption that the Clark's Completion of the input program is true, we then use a downward pass to propagate that the root node of the LNN has lower bound 1.0 and upper bound 1.0;
- finally, on the resulting LNN, we perform the Upward–Downward algorithm until the truth bounds of all subformulae converge, which happens in a finite number of steps.
For now, this is a basic hobby project and by no means considered to be stable. For more details on certain intricacies, feel free to issue a question.