Efficiently solving planning, management, or operational control problems requires advanced optimization algorithms. These algorithms use models to explore a range of alternative solutions and to evaluate the potential impact of proposed solutions within a realistic simulated environment. The generation and assessment of these solutions are meticulously organized and managed by systems employing diverse algorithmic approaches.
Area leaders:
Intelligent control for networked systems - IDSIA NEW
Scientific area
Intelligent control for networked systems
SUPSI Image Focus
Graph-based reinforcement learning in structured environments
The research investigates methodological advances for exploiting graph- based neural representations in reinforcement learning. The goal is to design a theoretical framework to learn composable policies i.e., a computational unit, a node-level policy, that takes part in the decision-making mechanism by processing information in an integrated form with respect to a committee of policies, possibly sharing the same parameters and exchanging messages along some edges. Thanks to this modular approach, the structure of the underlying system can be exploited, enabling the reuse of the same modules across different tasks.
Key research tasks:
- Theoretical understanding of the expressiveness of composable policies based on message passing. Existing methods for designing composable policies do not have theoretical justification or guarantees as such it is necessary to correctly frame the problem to ground design choices in theory.
- Spatio-temporal abstraction with hierarchical representations. Learning local actuator-level policies arguably makes spatio-temporal abstraction more difficult. We tackle this problem by introducing hierarchical, multi-layered, graph representations that naturally implement hierarchical reinforcement learning architectures.
- Control of metamorphic agents. Most existing methods focus on settings where the structure of the agents is fixed and does not change with time. The research allows agents to autonomously change their morphology and devise algorithms for automatic design and graph learning.
Combinatorial optimization and AI for intelligent planning and control
The design, planning and management of networked systems imply complex decision and control problems. The sinergy between combinatorial optimization and AI enables scalable and more accurate solution methods, driven by emerging challenges such as industrial electrification, digitalization and sustainability.
From a methodological perspective, AI supports the automated configuration of optimization algorithms, drives the exploration of search processes, and identifies promising solutions. Machine learning facilitates the decomposition of large-scale problems, while AI-based regression techniques efficiently approximate complex non-linear phenomena, enabling derivative-free optimization methods. Moreover, combinatorial optimization - through convex, robust, and mixed-integer formulations - offers principled and reliable tools for improving AI training and parameter tuning beyond traditional heuristic methods.
From a methodological perspective, AI supports the automated configuration of optimization algorithms, drives the exploration of search processes, and identifies promising solutions. Machine learning facilitates the decomposition of large-scale problems, while AI-based regression techniques efficiently approximate complex non-linear phenomena, enabling derivative-free optimization methods. Moreover, combinatorial optimization - through convex, robust, and mixed-integer formulations - offers principled and reliable tools for improving AI training and parameter tuning beyond traditional heuristic methods.
Optimal control and self-tuning of industrial machines
In recent years, IDSIA has developed active learning algorithms for the experimental design of real-time optimal controllers and for the automatic calibration of industrial machines. These methods enable the efficient optimization of complex parameters, significantly reducing time and costs compared to traditional manual trial-and-error procedures.
Beyond standard active learning techniques, such as Bayesian Optimization, IDSIA has introduced novel approaches based on preference learning, where decision-makers are iteratively asked to express simple qualitative pairwise preferences (e.g., “this is better than that”) between candidate decision vectors. This approach is particularly effective when the objective function cannot be easily quantified, either because it is qualitative in nature or because it involves multiple objectives with undefined priorities. By leveraging the human ability to compare alternatives rather than assign absolute scores, these methods include a new formulation of *Preferential Bayesian Optimization* based on skew-Gaussian processes, which has demonstrated superior performance compared to state-of-the-art preference-based Bayesian optimization methods.
These methodologies have been successfully applied to robotic sealing and assembly tasks, hyperparameter calibration for efficient implementations of model predictive control laws (including embedded systems), and several applied projects funded by Innosuisse and the European Union. Application areas include high-power laser cutting (in collaboration with Bystronic AG), drilling and wire electrical discharge machining (with Georg Fischer Ltd), and ultrashort-pulse laser systems for high-precision manufacturing and processing (with CSEM, BHF, and FEMTOPrint SA).
Beyond standard active learning techniques, such as Bayesian Optimization, IDSIA has introduced novel approaches based on preference learning, where decision-makers are iteratively asked to express simple qualitative pairwise preferences (e.g., “this is better than that”) between candidate decision vectors. This approach is particularly effective when the objective function cannot be easily quantified, either because it is qualitative in nature or because it involves multiple objectives with undefined priorities. By leveraging the human ability to compare alternatives rather than assign absolute scores, these methods include a new formulation of *Preferential Bayesian Optimization* based on skew-Gaussian processes, which has demonstrated superior performance compared to state-of-the-art preference-based Bayesian optimization methods.
These methodologies have been successfully applied to robotic sealing and assembly tasks, hyperparameter calibration for efficient implementations of model predictive control laws (including embedded systems), and several applied projects funded by Innosuisse and the European Union. Application areas include high-power laser cutting (in collaboration with Bystronic AG), drilling and wire electrical discharge machining (with Georg Fischer Ltd), and ultrashort-pulse laser systems for high-precision manufacturing and processing (with CSEM, BHF, and FEMTOPrint SA).
Relational spatio-temporal representations for prediction and control
This research line focuses on advancing relational representation learning for the prediction and control of complex systems, where both spatial and temporal information are essential.
Key research challenges:
Key research challenges:
- Designing learning agents with architectural inductive biases that enable the acquisition of spatio-temporal representations through sequential interaction with the environment in model-free reinforcement learning.
- Developing relational spatio-temporal representations for time-series analysis, with applications to forecasting and imputation in high-dimensional multivariate time series. In this context, spatio-temporal graph neural processing frameworks are extended to handle generic multivariate time series, such as those generated by sensor networks.
- Integrating these approaches within model-based reinforcement learning, where policy learning is supported by a learned spatio-temporal, graph-based model of the environment.
Computational biophysics
IDSIA hosts a Computational Biophysics Unit (CBU) that applies a wide spectrum of molecular and multiscale computational techniques to study complex biological systems. The unit’s research focuses on areas such as drug delivery systems and nanoparticle design and optimization, in silico structure- and ligand-based virtual screening, the investigation of drug mechanisms of action, and the self-assembly of biopolymers.