D6.17: QRBS Evaluation Report

This report is the last deliverable of use case 6 – Quantum Rule-Based Systems (QRBS) for breast cancer detection - of the NEASQC project. It presents the results of the software evaluation and benchmarking against classical solutions. The evaluation was done in five steps: Check the programmability of the developed library, i.e., the possibility of developing applications by other programmers. Integration of the library into Eviden Qaptiva environment. Test the proposed algorithms using a basic synthetic case. Benchmark the solution against classical standard for Invasive Ductal Carcinoma (IDC) diagnosis. Evaluate the solution on noisy emulated Quantum Processing Units (QPU). As a consequence of this work, the QRBS software release was extended with four tutorials using notebooks, some issues that limited the usability were identified and corrected, and additional support for emulators with larger capacity and other capabilities (such as the possibility of emulating noise using Qaptiva software) was introduced. The main conclusions extracted from the evaluation are: The results obtained for all the cases in a noiseless environment are those expected for the defined problems. These values present a high degree of similarity with those of classical models, as shown by the comparison of QRBS with fuzzy logic. In fact, the problem is not inferential; the problem is derived from the classification criteria. When comparing classical and quantum methods, it can be observed that the corresponding results are consistent. However, the QRBS approach improves the outputs adding intrinsic uncertainties to the inferences, which are not available in the classical approach. The results obtained show an almost identical behaviour among them, thus reinforcing the hypothesis (in both, qualitatively and quantitatively perspectives) that Quantum Computing is a more general paradigm, with the capacity to encompass the different classical models of symbolic reasoning. Noise affects the results catastrophically when is higher than a threshold, producing for the current QPUs useless results. However, for the tested models, a moderate improvement in the quality of the QPUs could produce acceptable results. Of course, the executed models do not justify yet the usage of real QPUs, because they are still small, allowing the classical ideal emulation. Experiments with larger models are needed to assess clearly the effect of the noise.