Millán Ruiz, David2024-05-202024-05-202010-07https://hdl.handle.net/20.500.14468/14127This work describes a novel approach to workforce distribution in dynamic multi-agent systems based on backboard architectures. These environments entail quick adaptations to a changing environment that only fast greedy heuristics can handle. These greedy heuristics consist of a continuous re-planning, considering the current state of the system. As these decisions are greedily taken, the workforce distribution may be poor for middle and/or long term planning due to incessant wrong movements. The use of parallel memetic algorithms, which are more complex than classical, ad-hoc heuristics, can guide us towards more accurate solutions. In order to apply parallel memetic algorithms to such a dynamic environment, we propose a reformulation of the traditional problem, which combines predictions of future situations with a precise search mechanism, by enlarging or diminishing the timeframe considered. The size of the time-frame depends upon the dynamism of the system (smaller when there is high dynamism and larger when there is low dynamism). This work demonstrates how nearly optimal solutions each v seconds (size of the time-frame) outperforms continuous bad distributions when the right size of the time-frame is determined, and predictions and optimisations are properly carried out. Specifically, we propose a neural network for predicting future system variables and a parallel memetic algorithm to perform the assignment of incoming tasks to the right agents, which outperforms other conventional approaches. Additionally, we propose a modification of the resilient back-propagation algorithm and evolutionary operators based on meta-heuristics. To conclude, we test out our method on a real-world production environment from Telefónica which is a large multinational telephone operator.eninfo:eu-repo/semantics/openAccesstesis de maestría