Several forms of model-based evolutionary design and optimization have been explored in PACE, including:
- the Model-based Genetic Algorithms-design (MGA-design), - the Predictive Neural Nets-design (PNN- design) - the Particle Swarm Optimization-design (PSO-design).
The MBE-design developed in PACE is conceived as an evolutionary process (instead of an apriori choice), where the selection of all the parameters involved (frequently hundreds of parameters) is the result of a sequential procedure where an initial small design is transformed by the information provided by real experiments. This information is captured by statistical models and introduced in the subsequent design of the evolutionary path. The design is then the result of an interactive process where each generation of real experiments, through modelling, can choose the next generation realizing an efficient and effective search of the space.
In Model-based Genetic Algorithms-designs (MGA-design) an initial random (and small) population of experiments is conducted and the results analyzed and modelled. A large family of statistical models are considered, such as regression models with different probability distributions and evolvable structure. From modelling it is possible to obtain information on the role of the mixture components and their interaction on the experimental result, and also on the effect of the several laboratory protocols. The evolution of the approach is then determined by a genetic algorithm, where the construction of the following generation of experiments is governed by the paradigm of evolution and the information achieved from modelling. (Forlin, et al., 2007).
In Predictive Neural Nets-designs (PNN- design) a stochastic neural network model is built on a very small set of experiments (random set), and then used to predict the results of all the experiments in the unknown space. The best experiments, according to a predefined optimality criterion, are then chosen to increase the initial set of experiments, on the results of which one then builds a different neural network model. This model learns from the new set of data and develops more accurate predictions on the rest of the space. The procedure continues till a convergence optimality value is reached. The approach is adaptive, since the model changes with the increasing set of experiments, and evolutionary, since the search strategy evolves with the information achieved at each step of the procedure.
In Particle Swarm Optimization-design (PSO-design) each experiment is regarded as a “particle” in the search space (particle of a swarm), which adjusts its position on the way to the optimum according to its own “flying” experience and the “flying” experience of other particles. The design evolve because each particle moves in the search space with an adaptable velocity based on information coming from the particle itself (the experimental result), as well as from the rest of the particles. Also this design is evolutionary, adaptive, and based on information and communication of information. (Forlin, et al. 2007).