How do humans master a new skill? They draw on prior experience, adapt quickly, and build from simple to complex. Replicating this capacity in artificial agents is one of the central challenges of modern reinforcement learning. The idea is to develop a hybrid framework combining meta-learning and continual learning in a two-phase pipeline. In the first phase, a meta-learner is trained on a broad distribution of locomotion tasks, learning a warm-start initialization that generalizes fast. In the second phase, this initialization seeds a continual learner that adapts to a sequence of new tasks without forgetting prior knowledge. The framework features a dual-objective training scheme and a sparse masking mechanism that partitions the network into a shared backbone and per-task private subnetworks (e.g., Hebbian-inspired). Tasks are presented in curriculum order, from dynamically simpler to progressively harder environments. The goal is to investigate whether such a unified framework can effectively transfer knowledge across tasks of increasing difficulty, and to explore open questions such as: how can a meta-learning algorithm provide the best warm-start for continual adaptation, whether the curriculum ordering can be learned rather than hand-designed, and how different sparse masking strategies compare in preserving and reusing past knowledge.
40% theory, 40% software, 20% testing
Requirements: Python, reinforcement learning
Contacts: Erik Nielsen, Giovanni Iacca
A Digital Twin (DT) model consists of physical entities, virtual counterparts, and their interactions. It uses descriptive, predictive, and prescriptive analytics for decision-making support. The Cognitive Digital Twin (CDT) enhances DTs with cognitive capabilities by integrating semantic technologies such as ontologies and knowledge graphs. This enables human-like intelligence, including attention, reasoning, and decision-making, allowing the CDT to recognize and adapt to complex behaviors dynamically. Advancing even further, the Intelligent Digital Twin (iDT) merges AI with human decision-making, fostering an AI-human collaborative environment. Despite their potential, DTs, CDTs, and iDTs are still emerging concepts, with limited real-world implementation. Most studies are theoretical, focusing on the envisioned capabilities of these technologies. This project aims to develop a digital twin use case for a hypothetical supply chain, demonstrating its descriptive, predictive, and prescriptive capabilities.
40% theory, 40% software, 20% testing
Requirements: Python
Contacts: Stefano Genetti, Giovanni Iacca
In Supply Chain Management (SCM), Digital Twin (DT) technology, particularly Intelligent DTs (iDTs), has emerged as a powerful tool, integrating human expertise with AI to enhance decision-making. However, most AI implementations in SCM rely on black-box models, lacking interpretability and facing resistance from practitioners who demand transparency. To address this, we developed Interactive Evolutionary Learning Decision Trees (IELDT), an Interpretable AI (IAI) framework that combines Grammatical Evolution (GE) and Reinforcement Learning (RL) to optimize production scheduling. Through an iterative optimization-simulation loop, IELDT refines decision tree scheduling policies using digital twin simulations. At each iteration, a web application presents the top two decision trees, together with natural language explanations from LLMs, allowing users to vote for their preferred policy. This fosters a human-in-the-loop framework, combining AI with human expertise to guide the algorithm toward user-preferred and better solutions. From this starting point, this project opens several potential directions: enhancing user interaction by allowing modifications to decision tree values and structure; supporting multiple users with independent web access to enable expert collaboration; enabling user-input examples of natural language descriptions rather than fixed prompt engineering to refine LLM-generated explanations, reducing configuration complexity and improving interpretability; study how to evaluate LLMs applying recent quality metrics; expanding the methodology beyond decision trees to explore alternative interpretable models.
30% theory, 40% software, 30% testing
Requirements: Python
Contacts: Stefano Genetti, Giovanni Iacca
Demand forecasting involves predicting future customer demand, enabling businesses to optimize production, inventory, and resource allocation. Recent advancements highlight the potential of AI in enhancing demand forecasting. However, most of the AI models are "black boxes", meaning that their internal workings are neither visible nor interpretable, and they fail to provide explanations for their predictions. This lack of transparency hinders their widespread adoption in industry where stakeholders, especially in business contexts, demand an understanding of the causal relationships and rationales behind model predictions. Explainable Artificial Intelligence (XAI) aims to bridge the gap between AI accuracy and transparency by offering insights into how and why models make decisions. This project aims to evaluate XAI solutions in demand forecasting, comparing them to black-box models to balance accuracy with transparency, fostering greater trust in AI-driven forecasting in industry.
20% theory, 40% software, 40% testing
Requirements: Python
Contacts: Stefano Genetti, Giovanni Iacca
This project explores Multi-Agent Systems (MAS) as a novel approach to supply chain management (SCM). MAS consist of autonomous, intelligent agents that perceive their environment, make decisions, and act to optimize operations. In SCM, they can manage supplier negotiations, client orders, inventory, and logistics, dynamically adapting to demand and supply fluctuations. Key applications include supplier bid optimization, real-time logistics monitoring, and adaptive decision-making. The project could compare MAS with traditional optimization methods, explore agent communication protocols, and study knowledge representation. Additionally, integrating Large Language Models (LLMs) could enhance agent reasoning, enabling chain-of-thought decision-making and problem-solving. Finally, we could examine human feedback integration to create a more collaborative MAS ecosystem.
30% theory, 60% software, 10% testing
Requirements: Python
Contacts: Stefano Genetti, Giovanni Iacca
The Job Shop Scheduling Problem is a widely studied NP-hard, industrially relevant problem. For large instances of the problem, the MILP formulations, solved with state-of-the-art open source optimizers such as CP-SAT, can be computationally demanding in terms of memory and time. In recent years, the research has moved toward hybridizing exact methods and meta-heuristics to solve this kind of problem. One exact method that could lead to optimal solutions is encoding it as a Satisfiability Modulo Theory (SMT) and solving it with an SMT solver like OptiMathSMT or Z3. This project aims at refining the encoding of the JSSP as an SMT problem, to use the maxSMT solver, which performs better than OptiMathSMT, and get feasible solutions for large instances of the JSSP. The project continues by optimizing the feasible solutions found by using Genetic Algorithms, Local Search, Variable Neighborhood Search, Simulated Annealing, or other search methods. As a final step, the comparison between the existing hybrid approach with CP-SAT + GA and the hybrid approach developed (SMT + search method) is requested.
30% theory, 40% software, 30% testing
Requirements: Python, SAT solving
References:
Contacts: Chiara Camilla Rambaldi Migliore, Marco Roveri, Giovanni Iacca
Many embedded applications have strict energy, memory, and time constraints, making neural network (NN) inference particularly challenging. Input-adaptive models like MoEs (mixture of experts, see, e.g., Fast-Inf) and FFFs (Fast Feed-Forward networks) exploit the divide-and-conquer approach to achieve efficient inference with extremely lightweight computational demands and minimal latency. However, with numerous hyperparameters influencing their architecture, optimizing training strategies and identifying the most effective approach to leverage their potential across different architectures remain significant challenges.
Some possible steps are:
Analyzing different configurations of FFF networks, including varying the number of leaves, different widths and depths, and using possible regularization techniques.
Benchmarking these configurations across multiple tasks such as simple image classification (MNIST), Keyword Spotting (KWS), and Human Activity Recognition (HAR).
Comparing model performance, efficiency, and complexity.
If time allows, extending the study to mixed models, MoE with FFF-based experts, or FFF with MoE-based leaves.
20% theory, 50% software, 30% testing
Requirements: Python
Contacts: Renan Beran Kilic, Giovanni Iacca
Many embedded applications have strict energy, memory, and time constraints, making neural network (NN) inference particularly challenging. Input-adaptive models like MoEs (mixture of experts, see, e.g., Fast-Inf) and FFFs (Fast Feed-Forward networks) exploit the divide-and-conquer approach to achieve efficient inference with extremely lightweight computational demands and minimal latency. However, understanding how nodes/routers make input decisions as they route information toward the leaves/experts and how these decisions affect leaves/experts remains a significant challenge.
Some possible steps are:
Investigating the interpretability of input-adaptive models, focusing on: Fast Feedforward (FFF) network, Mixture of Experts (MoEs).
Applying explainability techniques to understand model decisions of the model nodes (for FFFs), router (MoEs), or leaves/experts.
Comparing interpretability across multiple tasks such as simple image classification (MNIST), Keyword Spotting (KWS), and Human Activity Recognition (HAR).
Analyzing how model adaptation affects interpretability and decision transparency.
If time allows, extending the study to mixed models, MoE with FFF-based experts or FFF with MoE-based leaves
30% theory, 40% software, 30% testing
Requirements: Python
Contacts: Renan Beran Kilic, Giovanni Iacca