Thesis Proposals
Send an email to livia.lestingi@polimi.it specifying:
the proposal(s) you are interested in
your current grade average
your current career progress (i.e., number of exams yet to be passed)
It is fine if you do not possess the listed skills when starting out, but do factor in the extra time needed to acquire them.
All listed projects, if completed, are eligible for presentation as a thesis with discussant (i.e., "tesi"). Requests to work on a thesis without a discussant (i.e., "tesina") will be evaluated at the discretion of the supervisor.
Learning Models of Human Behavior During Mass Emergencies
Rescue services increasingly exploit autonomous systems as a support tool during mass emergencies.
For example, fleets of drones can be deployed to monitor an enclosed area and interact with human
subjects to foster cooperation and ease the evacuation process.
As a highly safety-critical setting, ensuring the drones' decision-making policies guarantee certain
properties is fundamental to maximizing life-savings. To this end, the Open University
and Politecnico di Milano are developing a formal game-theoretic approach to synthesizing strategies
for drones deployed on the scene.
This thesis aims to explore how learning models of human behavior at real-time (i.e., during the emergency) and incorporating them into the formal approach improves the efficiency of drones' decisions. To this end automata learning algorithms will be employed. The developed approach will be tested in simulation.
References: 10.1145/3654439
Inspiration: 10.1007/s00165-020-00509-0
Skills: Large Language Models; Natural Language Processing; Automated Software Testing.
Preferred M.Sc. Programs: CS.
Est. Effort: 8(+/-)2 months
This thesis aims to explore how learning models of human behavior at real-time (i.e., during the emergency) and incorporating them into the formal approach improves the efficiency of drones' decisions. To this end automata learning algorithms will be employed. The developed approach will be tested in simulation.
References: 10.1145/3654439
Inspiration: 10.1007/s00165-020-00509-0
Skills: Large Language Models; Natural Language Processing; Automated Software Testing.
Preferred M.Sc. Programs: CS.
Est. Effort: 8(+/-)2 months
Policy Learning in Automated Digital Twins
Digital Twin (DT) development is a potentially disruptive technology to the manufacturing field,
allowing for an efficient and powerful real-time analysis of complex real-life systems.
However, creating and maintaining DTs requires huge manual effort. Therefore, the
Auto-Twin project coordinated by Politecnico di Milano
aims at developing a framework to automatically generate and update DTs.
Decisional policies employed by the system (e.g., the switching rule between alternative paths for an item through the plant) are pivotal to effective DTs but often opaque and complex to learn. In this perspective, automata learning and process mining are employed to learn mathematical models of the system under analysis from field-collected data. This project aims to extend the employed model learning techniques to learn decision policies. The developed methodology will be tested on toy plants and real-life use cases from the industrial practice.
References: 10.1109/MIS.2022.3215698
Skills: Digital Twins; Automata Learning; Formal Modeling; Process Mining.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
Decisional policies employed by the system (e.g., the switching rule between alternative paths for an item through the plant) are pivotal to effective DTs but often opaque and complex to learn. In this perspective, automata learning and process mining are employed to learn mathematical models of the system under analysis from field-collected data. This project aims to extend the employed model learning techniques to learn decision policies. The developed methodology will be tested on toy plants and real-life use cases from the industrial practice.
References: 10.1109/MIS.2022.3215698
Skills: Digital Twins; Automata Learning; Formal Modeling; Process Mining.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
Repairing Automatically Learned Models of Cyber-Physical Systems
Politecnico di Milano is developing an automata Learning algorithm that, fed with traces recording the behavior
of a Cyber-Physical System, mines a Stochastic Hybrid Automaton modeling the system under learning (SUL).
Given its data-driven nature, the resulting model only captures a subset of the possible SUL behaviors.
In reality, numerous variations of training traces may occur that are not captured by the learned model.
However, repeating the learning with a more extensive training dataset may demand significant time and resources.
Possible trace variations have been classified (e.g., the repetition of a known event or the occurrence of two known events in a different order). This thesis aims to develop automated corrections to the learned model to make it robust to trace variations without repeating the learning. The envisaged repair solutions shall be implemented and tested on toy use cases and real-life systems to assess the added value of a more extensive model and the loss of accuracy compared to re-training.
References: 10.1109/MIS.2022.3215698
Inspiration: 10.1007/978-3-319-21690-4_32, 10.1016/0890-5401(87)90052-6
Skills: Large Language Models; Natural Language Processing; Automated Software Testing.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
Possible trace variations have been classified (e.g., the repetition of a known event or the occurrence of two known events in a different order). This thesis aims to develop automated corrections to the learned model to make it robust to trace variations without repeating the learning. The envisaged repair solutions shall be implemented and tested on toy use cases and real-life systems to assess the added value of a more extensive model and the loss of accuracy compared to re-training.
References: 10.1109/MIS.2022.3215698
Inspiration: 10.1007/978-3-319-21690-4_32, 10.1016/0890-5401(87)90052-6
Skills: Large Language Models; Natural Language Processing; Automated Software Testing.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
Proactive Self-adaptation of Assistive Robot Controllers
Polimi has developed a framework for the development of robotic applications in service settings (e.g., hospitals, home assistance, education).
The framework includes a design-time analysis phase that supports the specification of human-machine teaming missions.
Given the specification, requirements for the mission can be proven (e.g., probability of success greater than 0.9,
maximum fatigue of human agents less than 0.5).
However, a number of factors can dynamically change and affect the satisfaction of such requirements. These factors include quantities observable from the external environment as well as configuration parameters of the system that may be controlled at runtime.
This thesis project aims at making robotic applications self-adaptive, that is, able to adjust its controllable parameters automatically at runtime with the goal of maximizing the likelihood of satisfying all the requirements. Self-adaptation shall be proactive, meaning that the system shall be able to forecast possible violations and adapt itself to avoid them. The selected adaptation measure shall be subject to verification to assess its soundness, ensuring the verification result is available within the estimated remaining duration of the mission.
The thesis shall result in a software tool that support the proactive self-adaptation as well as an empirical evaluation of feasibility, and cost-effectiveness using one or more case studies.
References: 10.1016/j.robot.2023.104387
Inspiration: 10.1145/3524844.3528056
Skills: Automata Theory and Formal Modeling; Multiobjective Optimization; Machine Learning.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
However, a number of factors can dynamically change and affect the satisfaction of such requirements. These factors include quantities observable from the external environment as well as configuration parameters of the system that may be controlled at runtime.
This thesis project aims at making robotic applications self-adaptive, that is, able to adjust its controllable parameters automatically at runtime with the goal of maximizing the likelihood of satisfying all the requirements. Self-adaptation shall be proactive, meaning that the system shall be able to forecast possible violations and adapt itself to avoid them. The selected adaptation measure shall be subject to verification to assess its soundness, ensuring the verification result is available within the estimated remaining duration of the mission.
The thesis shall result in a software tool that support the proactive self-adaptation as well as an empirical evaluation of feasibility, and cost-effectiveness using one or more case studies.
References: 10.1016/j.robot.2023.104387
Inspiration: 10.1145/3524844.3528056
Skills: Automata Theory and Formal Modeling; Multiobjective Optimization; Machine Learning.
Preferred M.Sc. Programs: CS, ATM.
Est. Effort: 8(+/-)2 months
Towards AI-based Oracles for Automata Learning
Active automata learning is a long-standing research area targeting the synthesis of minimal Deterministic Finite-state Automata (DFA).
Foundational learning algorithms (e.g., L*) and tools (e.g., the LearnLib library) can be found in the literature.
Well-established techniques rely on the assumption that an omniscient oracle of the System Under Learning (SUL) is available. However, as complex Cyber-Physical systems grow increasingly widespread, it is paramount to challenge the assumption that perfect knowledge about the SUL is feasible in practice. Indeed, novel data-driven automata learning techniques targeting Hybrid Automata or Markov Decision Processes have been recently introduced.
With this thesis proposal, we foster the investigation of the following research issues:
Inspiration: 10.1109/EMSOFT.2015.7318273; 10.1007/978-3-030-30942-8_38
Skills: Automata Theory; Automata Learning; Machine Learning Models.
Preferred M.Sc. Programs: CS.
Est. Effort: 8(+/-)2 months
Well-established techniques rely on the assumption that an omniscient oracle of the System Under Learning (SUL) is available. However, as complex Cyber-Physical systems grow increasingly widespread, it is paramount to challenge the assumption that perfect knowledge about the SUL is feasible in practice. Indeed, novel data-driven automata learning techniques targeting Hybrid Automata or Markov Decision Processes have been recently introduced.
With this thesis proposal, we foster the investigation of the following research issues:
- How do well-established automata learning algorithms (e.g., L*) perform in a realistic scenario, thus with a non-omniscient oracle relying on partial observations of the SUL?
- How can these algorithms be modified/extended to work in a more realistic setup?
- To this end, can an AI-based oracle (e.g., a Neural Network) be exploited?
- Could this setup simultaneously serve the purpose of generating an explainable model of the AI-based oracle?
Inspiration: 10.1109/EMSOFT.2015.7318273; 10.1007/978-3-030-30942-8_38
Skills: Automata Theory; Automata Learning; Machine Learning Models.
Preferred M.Sc. Programs: CS.
Est. Effort: 8(+/-)2 months
