Optimisation is one of the most important and challenging part of any engineering design. In real world design problems one faces multiobjective optimisation under constraints. The optimal solution in these cases is not unique because the objectives can contradict each other. In such cases, a set of optimal solutions which forms a Pareto frontier in the objective space is considered. There are many algorithms to generate the Pareto frontier. However, only a few of them are potentially capable of providing an evenly distributed set of the solutions. Such a property is especially important in real-life design because a decision maker is usually able to analyse only a very limited quantity of solutions. This thesis consists of two main parts. At first, it develops and gives the detailed description of two different algorithms that are able to generate an evenly distributed Pareto set in a general formulation. One is a classical approach and called Directed Search Domain (DSD) and the other, the cylindrical constraint evolutionary algorithm (CCEA), is a hybrid population based method. The efficiency of the algorithms are demonstrated by a number of challenging test cases and the comparisons with the results of the other existing methods. It is shown that the proposed methods are successful in generating the Pareto solutions even when some existing methods fail. In real world design problems, deterministic approaches cannot provide a reliable solution as in the event of uncertainty, deterministic optimal solution would be infeasible in many instances. Therefore a solution less sensitive to problem perturbation is desirable. This leads to the robust solution which is the focus of the second part of the thesis. In the literature, there are some techniques tailored for robust optimisation. However, most of them are either computationally expensive or do not systematically articulate the designer preferences into a robust solution. In this thesis, by introducing a measure for robustness in multiobjective context, a tunable robust function (TRF) is presented. Including the TRF in the problem formulation, it is demonstrated that the desirable robust solution based on designer preferences can be obtained. This not only provides the robust solution but also gives a control over the robustness level. The method is efficient as it only increases the dimension of the problem by one irrespective of the dimension of the original problem.

[pt] Nesse trabalho, propomos o problema AdInvest, que modela o processo decisiório de alocação de investimento em marketing digital do ponto de vista do anunciante. Para o problema proposto, definimos um algoritmo chamado balGreedy, e provamos suas garantias para instâncias determísticas e estocásticas do AdInvest. Os teoremas provados garantem ao nosso algoritmo resultados de pior caso relativamente próximos ao OPT, em diversos tipos de instâncias levantadas ao decorrer do trabalho. Em especial, focamos nas instâncias que modelam o efeito de saturação das audiências, que se faz presente na dinâmica de anúncios online. Como mostrado nos experimentos computacionais, o algoritmo balGreedy se mostrou consistentemente eficiente em comparação com as políticas alternativas adotadas, tanto nas instâncias que foram geradas por simulação, quanto em instâncias reais obtidas a partir de dados de um anunciante do Facebook Ads. / [en] In this work, we propose the problem AdInvest, which models the decision-making process for allocating investment in digital marketing from the advertiser perspective. For the proposed problem, we define an algorithm called balGreedy, and we prove its guarantees in deterministic and stochastic instances of the AdInvest. The proven theorems assure to our algorithm worst-case results relatively close to OPT, in several types of instances raised during the work. In particular, we focus on the instances that model the audience saturation effect, which is present in the dynamics of online advertisements. As shown in the computational experiments, the balGreedy algorithm had been consistently efficient compared to the alternative policies adopted, both in the instances generated by simulation and in real instances built from the data of a certain Facebook Ads advertiser.