Excellence

Main Objective

The main objective of this project is the development, testing, and validation of two optimization algorithms and related GUI.

The ﬁrst optimization algorithm is a qualitative optimization algorithm. Qualitative optimization takes as input qualitative feedback provided by the user [1]. Indeed, instead of using quantitative cost functions (based on quantitative measurements and on quantitative performance metrics) ascommon for quantitative optimization algorithms, such optimization algorithms will make use of user preferences to guide the optimization of atarget process. The proposed method aims for solving optimization problems in which the decision-maker (i.e., the user) cannot quantitativelyevaluate the objective function, but rather can only express a preference such as "this is better than that" between two candidate decision vectors(i.e., experiments). The proposed algorithm aims at reaching the global optimizer by iteratively proposing to the decision maker a newcomparison to make, based on actively learning a surrogate of the latent (unknown and perhaps unquantiﬁable) objective function from pastsampled decision vectors and pairwise preferences. In other words, in such a way, the user can compare two process outcomes resulting from two diﬀerent process settings. By providing the information on the preferred experiment to the algorithm (e.g., experiment 1 is better than experiment 2, the optimization method can update its internal modeling, providing a new setting to be tested (i.e., for which a new outcome of the process isobtained). This new outcome is compared with the best one so far, providing the algorithm with the new qualitative feedback (i.e., preference).This procedure is indeed iteratively repeated until a satisfactory process setting is achieved (i.e., the process settings are optimized and the maximum quality of its outcome is reached). SUPSI is an expert in this area, having developed the method in [1] and proposed improved(similar) methods [2] and having experience in applying these approaches to real industrial applications [3,4]. A preference- based optimizationtool has been also developed, tested, and evaluated in the SHOP4CF PROPHET project, where the optimization toolbox has been validated up to TRL 7.

While preference-based optimization is an eﬀective and powerful tool, it can be enhanced with the capabilities of quantitative optimization. In fact, the main limitation of preference-based optimization is related to the fact that the user's qualitative feedback has to be consistent during the optimization process. To mitigate this risk, quantitative objectives can be included, and a hybrid (i.e., fusing qualitative and quantitative optimizationfeatures) optimization algorithm can be developed. This will be the second optimization approach developed by this project. Bayesianoptimization and preference-based optimization will be fused into a single optimization algorithm to guarantee the maximum performance andﬂexibility of the optimization toolbox. In such a way, the novel optimization algorithm will make it possible to be employed in any situation,maximizing its exploitation in industrial contexts.

Both algorithms will be enhanced by transfer learning capabilities. This capability allows the algorithm to create global modeling for similarapplications, allowing it to speed up the tuning of similar processes to the ones already optimized. This is possible by exploiting the alreadygathered data to create the global process modeling under (partially) diﬀerent operative conditions, improving the ﬂexibility, adaptability, and generalizability of the proposed algorithms [5].

[1] Bemporad, Alberto, and Dario Piga. "Global optimization based on active preference learning with radial basis functions." Machine Learning 110 (2021): 417-448.
[2] Dao, Le Anh, et al. "Experience in Engineering Complex Systems: Active Preference Learning with Multiple Outcomes and Certainty Levels." arXiv preprint arXiv:2302.14630 (2023).
[3] Roveda, Loris, et al. "A human-centric framework for robotic task learning and optimization." Journal of Manufacturing Systems 67 (2023): 68-79.
[4] Maccarini, Marco, et al. "Preference-Based Optimization of a Human-Robot Collaborative Controller." IFAC-PapersOnLine 55.38(2022): 7-12.
[5] Roveda, Loris, et al. "Multi-Objects Robotic Grasping Optimization Employing a 2D camera." 2022 International Conference onElectrical, Computer, Communications and Mechatronics Engineering (ICECCME). IEEE, 2022.

Key Marketable Innovations

The outcome of the project will be a tool for preference-based optimization, i.e., an AI-based interface guiding the operators in processoptimization.

The resulting system will be ready to use in any manufacturing processing where qualitative optimization might be needed (i.e., where quantitative optimization criteria, measurements/sensors, and quantitative cost functions are not available). The developed tool will be general-purpose and customizable for any speciﬁc use.

The developed tool will be extensively tested and evaluated in the two use cases presented in the project, reaching a TRL 8-9, and being ready to go to the market for its exploitation.

In addition, a modiﬁed version of the algorithm will be developed to integrate both quantitative and qualitative optimization. In such a way, acombination of both quantitative and qualitative feedback will be possible, fusing two distinct optimization algorithms. In particular, consideringquantitative optimization, Bayesian optimization-based algorithms will be employed. A purely Bayesian optimization-based tool is alreadyavailable from SUPSI, and it has been developed, tested, and validated in a previous EIT-M project (IMPALA). The capabilities and potential ofthis tool will be enhanced with the ones of preference-based optimization, providing a hybrid quantitative/qualitative optimization tool that canbe exploited in any operative scenario.

Both algorithms will be enhanced by transfer learning capabilities. This capability allows the algorithm to create global modeling for similarapplications, allowing it to speed up the tuning of similar processes to the ones already optimized (i.e., reducing the required time, personnel,resources, and costs related to the optimization procedure to be executed - being such optimization procedure iterative and experiments-based). This is possible by exploiting the already gathered data to create the global process modeling under (partially) diﬀerent operative conditions,improving the ﬂexibility, adaptability, and generalizability of the proposed algorithms.

To highlight the potential of the proposed approaches, two use cases are considered in this project:

[Tec-Eurolab]: optimization of parameters of Industrial Computed Tomography scans;
[SMARTZAVOD]: optimization of polymer printing and automatic post-processing parameters for hybrid 3D printer.

These two applications, in diﬀerent domains, show the wide applicability and ﬂexibility of the proposed software components, allowing for the extensive validation of the proposed approaches.

The proposed solution terriﬁcally improves the state of the art in these industrial process optimization tasks. We expect the proposed user-centered optimization tool to drastically reduce the number of experimental trials to tune the machine parameters by providing more accurate parameters over time.

At the same time, the quality of the ﬁnal product is increased through continuous control of the system parameters, aﬀecting user satisfaction and reducing the costs related to scrap.

The operator can give value to its experience by providing feedback to the system without requiring any programming skills. In fact, the front end, designed taking into consideration the user needs, will provide a friendly interface to the system, reducing the engaging time and possible userresilience. The user will, indeed, focus on more valuable activities while the algorithms control the process. The barrier related to themanufacturing process is reduced by collecting knowledge from experienced operators and delivering it to less skilled operators.

A learning path will be produced within the project and can be used to introduce the main concepts of process optimization and train the end users and all the potential customers.

The marketable outcome of the project will, therefore, be TWO optimization TOOLBOXES. The ﬁrst will be a PREFERENCE-BASED optimization TOOLBOX, while the other will be a HYBRID (i.e., combining quantitative and qualitative optimizations) optimization TOOLBOX.These two software (including the GUI for simpliﬁed usage) will be available to the customers. In particular, the hybrid optimization software will provide a ﬂexible solution that can be employed in any scenario (with or without sensors, with or without quantitative metrics, etc.). Consideringthe novel nature of this hybrid algorithm, a patent for this technology will be considered.

The KEY PERFORMANCE TARGETS of the marketable innovation we can envision are:

Reduction of the time necessary for ﬁne tuning: min. 20%
Improvement of the ﬁne tuned system: around 10%

Ambitions and overlap beyond the state-of-the-art

Preference-based optimization is a powerful optimization tool to overcome limitations/issues related to quantitative optimization algorithms (i.e., lack of sensors, quantitative objective functions, etc.). Preference-based optimization algorithms can be found in the state of the art [1,2]. However, their application in real industrial applications is still far away to be widely found. SUPSI is one of the ﬁrst research centers to move this technology to the real industrial ﬁeld. Within the SHOP4CF PROPHET project, SUPSI applied this optimization method to a visual inspection system tuning application, making use of the experience of the expert user to ﬁne-tune a process. Indeed, the expertise of SUPSI in the ﬁeld, being one of the leaders and developers of these algorithms, is precious to develop, test, and validate such technologies in real industrial applications as proposed in this project, moving forward the state-of- the-art in the ﬁeld.

To maximize the performance of the optimization toolbox, a hybrid solution (i.e., combining qualitative and quantitative optimization capabilities) will be developed, tested, and evaluated. The hybrid toolbox will be developed by fusing Bayesian optimization and preference-based optimization. It has to be underlined that no existing algorithms can be found in the state of the art fusing these capabilities (i.e., qualitative and quantitative optimization). While diﬀerent approaches are describing Bayesian optimization [3], showcasing its potential [4,5,6], as for preference-based optimization [7,8], no approaches are found for this novel hybrid optimization algorithm concept. Therefore, the innovative and ambitious nature of the project is also in the development of advanced and novel algorithms that will be extensively tested in real operative conditions to reach TRL 8. A patent will be considered for this novel optimization algorithm.

Both algorithms will be enhanced by transfer learning capabilities. This capability allows the algorithm to create global modeling for similar applications, allowing it to speed up the tuning of similar processes to the ones already optimized. This is possible by exploiting the already gathered data to create the global process modeling under (partially) diﬀerent operative conditions, improving the ﬂexibility, adaptability, and generalizability of the proposed algorithms [9].

[1] Bemporad, Alberto, and Dario Piga. "Global optimization based on active preference learning with radial basis functions." Machine Learning 110 (2021): 417-448.
[2] Dao, Le Anh, et al. "Experience in Engineering Complex Systems: Active Preference Learning with Multiple Outcomes and Certainty Levels." arXiv preprint arXiv:2302.14630 (2023).
[3] Frazier, Peter I. "A tutorial on Bayesian optimization." arXiv preprint arXiv:1807.02811 (2018).
[4] Roveda, Loris, et al. "Human–robot collaboration in sensorless assembly task learning enhanced by uncertainties adaptation via Bayesian Optimization." Robotics and Autonomous Systems 136 (2021): 103711.
[5] Roveda, Loris, et al. "Enhancing object detection performance through sensor pose deﬁnition with bayesian optimization." 2021 IEEE international workshop on metrology for Industry 4.0 & IoT (MetroInd4. 0&IoT). IEEE, 2021.
[6] Roveda, Loris, Marco Forgione, and Dario Piga. "Robot control parameters auto-tuning in trajectory tracking applications." Control Engineering Practice 101 (2020): 104488.
[7] Roveda, Loris, et al. "A human-centric framework for robotic task learning and optimization." Journal of Manufacturing Systems 67 (2023): 68-79.
[8] Maccarini, Marco, et al. "Preference-Based Optimization of a Human-Robot Collaborative Controller." IFAC-PapersOnLine 55.38 (2022): 7-12.
[9] Roveda, Loris, et al. "Multi-Objects Robotic Grasping Optimization Employing a 2D camera." 2022 International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). IEEE, 2022.

Current Technology Readiness Level (TRL) and plan to reach TRL 8

The proposed preference-based optimization algorithm and software have been already validated in the SHOP4CF PROPHET project. In this project, the toolbox has been applied for the tuning of the parameters of an inspection system. More in detail, the inspection

system aims to identify defects in production parts. To correctly recognize a defect and its type, the inspection system software has to be tuned (i.e., properly setting some internal parameters related to the internal machine vision-based algorithms). Such a tuning has been performed making use of the proposed qualitative optimization toolbox. The user, iteratively, provides his/her judgment (i.e., preference) on a pair of experiments (the last setting suggested by the algorithm vs the best one so far). Thus judgment is based on the user's visual analysis of the defect recognition performance of the system with the applied setting. In such a way, an intuitive optimization procedure is implemented, only requiring the user to judge the performance of the system without requiring any manual tuning of it. The TRL reached by the project was 7. Indeed, to reach TRL 8, two end users have been included in this consortium, so that the validation of the toolbox can be performed in real applications, taking its usage to the production level.

SUPSI has already developed, evaluated, and tested quantitative optimization algorithms in previously funded projects (e.g., EIT-M IMPALA, Innosuisse HYPER, Innosuisse VIOLA II). The TRL of such optimization algorithms (based on Bayesian optimization) is 8. Indeed, by combining quantitative optimization methods together with qualitative optimization methods, the aim is to develop a novel optimization algorithm capable to overcome the weakness of such algorithms and improving their performance and ﬂexibility.

Therefore, SUPSI will integrate these two algorithms (with TRL 7 and 8 as described above), to reach a TRL 8 with this hybrid method. This will be possible based on the experience SUPSI already has with optimization methodologies, testing in real conditions, and integration of these algorithms in production plants (especially considering the previous EIT-M project IMPALA). To help SUPSI reach the standard for TRL 8, Santer Reply SpA will engineer the software, designing user-friendly and ad hoc GUIs to assist the user during the optimization process.

Gender balance, demographic diversity and inclusivity

The consortium acknowledges that there is a lack of an equitable distribution of female scientists/engineering at this moment, and recognizes the importance of gender equality in the planned activities. The companies participating in the consortium are committed to gender equality, equal pay and reconciliation measures. The involvement of women at all levels of management and research will also be promoted.

The project coordinator will take responsibility and take the necessary actions to avoid gender discrimination. Likewise, the participating organizations will support the involvement of female/other genders engineers and scientists in the activities conducted in the project.