Prof Abdelatif EL AFIA, Laboratoire des Smart Systèmes, Rabat IT Center. Full Professor à l’ENSIAS, à l’Université Mohamed V de Rabat. Il a obtenu son diplôme de Master, en 1993, en mathématiques appliquées à l’Université Sherbrooke en Canada, où il a eu son doctorat, en 1999, en recherche opérationnelle. Ses domaines de recherche incluent la programmation mathématique (stochastique et déterministe), les méta-heuristiques, l’apprentissage automatique, et plus particulièrement la performance des algorithmes d’apprentissage, ainsi que les algorithmes d’optimisation et leurs applications dans de nombreux domaines tels que : la chaîne d’approvisionnement, les prévisions et les systèmes de recommandation.
Talk: Risk Management of Particle Swarm Algorithm by Markov Model Control
A simulation model is generally an approximate representation of reality. Likewise, there are typically uncertainties concerning the association of the model and the actual system. Interpolation uncertainty comes from insufficient available data for computer model simulations and experimental measurements. Uncertainty can access mathematical models and experimental measurements in several contexts. Such as, parameter uncertainty is produced by the model parameters that are inputs to the mathematical model, but whose specific values are unknown and cannot be controlled or measured. Parameters are a key property to control algorithm’s exploration/exploitation capabilities. Hence, algorithmic uncertainty in general, derives from numerical errors and numerical approximations in the implementation of the computer model, also expressed as parametric variability, which indicates the uncertainty on the input variables of the model. For different input settings of simulation data or experimental measurements, the adequate setting is chosen as a prediction of the best corresponding responses. A quantitative risk model can be built based on historical achievements on an algorithm to calculate the impact of the uncertain parameters and the adjustments of each parameter that are made on defined outcomes. The process of risk analysis includes identifying and quantifying uncertainties, estimating their impact on algorithmic outcomes, building a risk analysis model that expresses these elements in a quantitative form, exploring the model through simulation, and making risk management decisions on algorithmic parameters that can help us to deal with risk.
The PSO is a well-known bioinspired algorithm applied to optimization problems, which involves a machine-learning technique inspired by birds freely flocking in search of food. In order to control the exploitation and exploration of PSO, there is a selection of PSO parameters that require to be tuned to have an efficient enhanced algorithm. A number of these parameters have been studied earlier from diverse aspects. Nevertheless, such parameters analyses have to be taken into consideration to set up an effective PSO tuned model. The primary intent behind parameters setting approaches of PSO is to incite complex global behaviors through local communications by sharing information between diverse agents and enhancing the learning capacity. Further, it may enable the swarm to adapt to unexpected variations (such as in the dynamic optimization) when they are interacting with more agents. The selection of the optimum parameter set can be based on an algorithmic model of the uncertainty of PSO based on a Markov chain. This approach can maximize the value of adaptive parameters selection rather than classical approaches which are based on the iteration number.
Markov model on PSO is used to constitute a generic model integrated as a performance enhancement of the particle swarm algorithm. PSO algorithm is improved by both offline and online configuration of its parameters based on a Markov chain on PSO. By integrating the identification of the suitable achievement level of the search process as a hidden state, PSO parameter adaptation will be more suitable if it is adjusted according to the classified PSO state.
The application of this model needs to analyze first the relationship between different parameter sets estimating the transition probability with the selection of the optimum parameter set in each state. This defined model of hidden PSO states and the method of classification and computing the related transition probabilities can be investigated for application in PSO for different parameters in the case of homogeneous or heterogeneous PSO structure.