1:30 - 2:00 Prof Thomas Parisini - "Distributed approximate uniform global minimum sharing"
"This lecture deals with the distributed minimum sharing problem: a set of decision-makers compute the minimum of some local quantities of interest in a distributed and decentralized way by exchanging information through a communication network. We propose an adjustable approximate solution which enjoys several properties of crucial importance in applications. Specific application contexts are illustrated first. The analysis shows that the proposed solution has good decentralization properties, and it is scalable in that the number of local variables does not grow with the size or topology of the communication network. Moreover, a global and uniform (both in the initial time and in the initial conditions) asymptotic stability result is provided towards a steady state which can be made arbitrarily close to the sought minimum. Exact asymptotic convergence can be recovered at the price of losing uniformity with respect to the initial time. A real industrial use case is described in the metal industry sector. ♦ slides ♦ [back to the event]
2:00 - 2:30 Dr Anastasia Borovykh - "On optimal datapoint selection for efficient learning"
"In this talk we will discuss several ideas around the distillation of a dataset for machine learning. The motivation for this may arise from different perspectives. For example, one may be interested to run a machine learning algorithm on a device with limited storage computing power. Alternatively, data may be expensive to acquire and understanding which datapoints are crucial for learning can help design acquisition strategies. Beginning with an overview of prior work, we will then discuss some preliminary results in using methods from mesh optimisation and function approximation and end with some directions for future work." [back to the event]
2:30 - 3:00 Dr Panos Parpas - "Recent developments in large scale optimisation"
"In this talk, we briefly review recent developments for optimizing large-scale models. We review work on stochastic algorithms, multilevel, and distributed methods. The talk's objective is to show how to take advantage of model structure and geometry to develop more efficient algorithms or solve otherwise intractable models." ♦ slides ♦ [back to the event]
3:00 - 3:30 Dr Chrysoula Kappatou - "On global optimization of data-driven models"
"Increase in data availability and digitalization of industrial processes widely promotes the use of data-driven models that are fast to evaluate. Yet, nonlinearity in the numerical representation of many operational and design problems requires global optimization methods for optimal decision making. Deterministic global optimization guarantees global solution in finite number of iterations. However, the computational cost associated with it largely limits its use to small problem sizes. Our open-source deterministic global optimization solver MAiNGO relying on McCormick relaxations and their subgradient propagation, enables reformulating the problems in a reduced-space formulation. Therefore, it allows for extensive computational savings. We present special theory and algorithm development for deterministic global dynamic optimization for a specific class of data-driven models, Hammerstein-Wiener models. We discuss the performance of our approaches using illustrative examples from process systems engineering applications. [back to the event]
4:00 - 4:30 Dr Stefan Vlaski - "Learning in heterogeneous networks"
"Distributed learning paradigms, such as federated and decentralised learning, have emerged as effective tools to harness vast amounts of data and computational resources available in dispersed locations to build powerful models in a privacy-preserving and communication efficient manner. Performance guarantees and gains of these approaches are well-established in homogeneous environments, where different agents in the network have similar objectives, capabilities and data distributions. In many practical applications, however, agents are interested in solving distinct albeit related tasks. In this talk we will review some new techniques and associated performance guarantees for designing intelligent networks in heterogeneous environments using tools from multi-task and meta-learning as well as social learning." ♦ slides ♦ [back to the event]
4:30 - 5:00 Dr Dario Paccagnan - "Robust labour planning in delivery networks"
"Labour planning concerns the scheduling of an organization's workforce to match its needs and demands. While for many business sectors achieving this goal is straightforward, labour planning is non-trivial whenever the demand is rapidly changing, or operations are carried out "just-in-time", such as in the logistics sector. In this context, it is paramount to design a labour plan that i) minimizes the organization's costs, ii) meets the demand, and iii) ensures that workers' shifts are as repetitive as possible. The latter requirement stems from the observation that fluctuations in the employees' schedules correlate strongly with high levels of absenteeism and turnover. Motivated by this observation, we tackle labour planning from a robust perspective in the context of one of the world largest delivery companies. Specifically, we propose tools and algorithms that leverage past demand realizations to design perfectly repetitive labour plans that minimize the organization's cost and meet the demand in all but a small fraction of future demand realization taken from the same (possibly unknown) distribution. Our approach builds upon the "Scenario Approach", thus providing computationally efficient algorithms and formal generalization guarantees to unseen demand realizations." [back to the event]
5:00 - 5:30 Prof Kin Leung and Zheyu Chen - "Optimisation by machine learning"
"Gradient-based iterative algorithms have been widely used to solve optimization problems, including resource sharing and network management. When system parameters change, it requires a new solution independent of the previous parameter settings considered by the methods. Therefore, we propose a learning approach that can quickly produce optimal solutions over a range of system parameters for constrained optimization problems. Two Coupled Long Short-Term Memory networks (CLSTMs) are proposed to find the optimal solution. Extensive numerical experiments using Alibaba datasets confirm the advantages of the new framework: (1) near-optimal solution can be obtained much quicker than gradient-based methods, (2) enhanced robustness as the CLSTMs can be trained using system parameters with distributions different from those used during inference to generate solutions." ♦ slides ♦ [back to the event]