Made available in DSpace on 2016-08-17T14:53:16Z (GMT). No. of bitstreams: 1 Leandro Rocha Lopes.pdf: 1075564 bytes, checksum: 01e184ed6d7c65323c0dfc1515da19a3 (MD5) Previous issue date: 2011-04-04 / Due to the increasing of technological development and its associated industrial applications, control design methods to attend high performance requests and reinforcement learning are been developed, not only, to solve new problems, as well as, to improve the performance of implemented controllers in the real systems. The reinforcement learning (RL) and discrete linear quadratic regulator (DLQR) approaches are connected by adaptive dynamic programming (ADP). This connection is oriented to the design of optimal controller for multivariable systems (MIMO). The proposed method for DLQR controllers tuning can been heuristic guidance for biased variations in weighting matrices of instantenous reward. The heuristics performance are evaluated in terms of convergence of heuristic dynamic programming (HDP) and action dependent (AD-HDP) algorithms. The algorithms and tuning are evaluated by the capability to map the plane-Z in MIMO dynamic system of third order. / Em decorrência do crescente desenvolvimento tecnológico e das consequentes aplicações industriais, técnicas de controle de alto desempenho e aprendizado por reforço estão sendo desenvolvidas não só para solucionar novos problemas, mas também para melhorar o desempenho de controladores já implementados em sistemas do mundo real. As abordagens do aprendizado por reforço e do regulador linear quadrático discreto (DLQR) são conectadas pelos métodos de programação dinâmica adaptativa. Esta união é orientada para o projeto de controladores ótimos em sistemas multivariáveis (MIMO). O método proposto para sintonia de controladores DLQR fornece diretrizes para construção de heurísticas polarizadas que são aplicadas na seleção das matrizes de ponderação da recompensa instantânea. Investiga-se o desempenho das heurísticas associadas com a sintonia de controladores lineares discretos e aspectos de convergência que estão relacionados com as variações QR nos algoritmos de programação dinâmica heurística (HDP) e Ação Dependente (ADHDP). Os algoritmos e a sintonia são avaliados pela capacidade em estabelecer a política de controle ótimo que mapeia o plano-Z em um sistema dinãmico multivariável de terceira ordem.

Having a accurate model of a system is essential for many applications today, especially those related to advanced process control. When executing a project often a lot of time is spent performing experiments on the real system to estimate a model. By designing higher quality experiments the time needed to estimate and identify these models can be reduced saving both resources and engineering efforts. This masters thesis investigates optimal input design to minimize the time needed to identify a linear time-invariant multivariable system fulfilling certain requirements on the model accuracy. Previous input designs mostly focused on sequential excitation but here the effects of using combined simultaneous and sequential excitation is investigated. The design is performed in simulations and evaluated in closed loop using a model predictive controller to further guarantee that the output constraints are not violated. The results indicate that there are many cases where using combined simultaneous and sequential excitation outperforms the previous methods. The effects of the color of the noise on the input design is investigated and the ability of different designs to estimate system delay is also studied. In addition it is shown how an iterative scheme can be used to guarantee that the accuracy requirements on the estimated model are met. / Att ha en god modell av ett system är viktigt för många applikationer idag, särskilt de som är relaterade till avancerad processtyrning. När man genomför ett projekt läggs ofta mycket tid på att utföra experiment på det verkliga systemet för att identifiera en modell. Genom att utforma experiment av hög kvalitet kan den tid som behövs för att identifiera dessa modeller minskas, vilket minimerar både processpåverkan och ingenjörsinsatsen. Denna masteruppsats undersöker metoder för optimal experimentdesign för att minimera tiden som behövs för att identifiera ett multivariabelt system där det finns krav på modellens noggrannhet. Tidigare metoder fokuserade mest på sekventiella experiment, men här undersöks effekterna av att använda en kombination av samtidiga och sekventiella experiment. Här används simuleringar som utvärderas i sluten loop med hjälp av en modellprediktiv regulator för att undvika att utsignalbegränsningarna inte överskrids. Resultatet indikerar att det finns många fall där användning av kombinerade samtidiga och sekventiella experiment överträffar tidigare metoder. Effekterna av färgat brus på ingångsdesignen undersöks och olika metoders förmåga att uppskatta systemfördröjning studeras också. Dessutom visas hur ett iterativt schema kan användas för att garantera att noggrannhetskraven på den uppskattade modellen uppfylls.