Computational networks & competition-based models solving complex causal interactions
Romdhane, Lofti Ben
In the current dissertation, we propose neural models to address the two"kinds" of causal reasoning: effect-to-cause and cause-to-effect. We propose a fuzzy clustering algorithm DD-[downside arrow]FLVQ, an enhanced version of [downside arrow]FLVQ, to address a variety of problems including effect-to-cause problems, pattern recognition, data compression. Experimental results reveals that DD-[downside arrow]FLVQ produces better classification than [downside arrow]FLVQ. To address cause-to-effect problems, computationally intractable problems, we propose essentially three neural models called respectively SOLVE, TOUNESUNIFIED and REASON. Our first model, SOLVE, is an extension of a recent abductive model to the cancellation clans of cause-to-effect reasoning. The second model, TOUNESUNIFIED, mechanizes the open, independent, incompatibility and monotonic classes and is based on an energy function. The application of TOUNESUNIFIED to artificially designed problems, as well as to real-world problems in the fields of legal reasoning, scientific theory formation and medical diagnosis reveals its [i.e. it's] the robustness and efficiency in handling complex causal problems. The third model, REASON, mechanizes causal problems in the open, independent and incompatibility classes. Again, REASON is based on an energy function and is mainly characterized by its efficient hardware implementation and its swiftness in obtaining optimal solutions even for large-scale problems. Finally, the major results and key features of the three proposals DD-[downside arrow]FLVQ, TOUNESUNIFIED and REASON are analyzed, and the limitations and future extensions of the research are discussed.
- Sciences – Thèses