(204y) Optimizing a Chemical Process With Bayesian Network (a new approach to optimization)

In this work Bayesian Networks are used as an optimization tool. Bayesian Network is an artificial intelligence tool that is used for different purposes like decision makings. To our knowledge, it is the first time that Bayesian Networks are developed for this purpose. Other optimization tools find optimum point of the target variable and show at what input variables values, this optimum point is achieved. On the other hand, in our work with Bayesian Networks , the target function is a function with variables mean and variance of the target variable, which means the value of the target function and distribution over this mean are get involved. As a result, stability of this variable over the optimum point take into consideration. In other words, Optimization algorithms optimize the target function with mathematical solutions and they do not care if this point has stability in the process or it happens only in a particular case by setting parameters on exact values that is not achievable in real processes. The procedure in our work is explained here: First of all a Bayesian Network is constructed with given data. In this work k-2 algorithm is used for structure learning. After that, the parameters of the network are learned and Conditional Probability Distributions (CPDs) of network variables are made. Then, some values (evidence) are assigned to independent nodes and probability flow in network is calculated. For the target variable a score function is defined. This function might be a linear combination of new calculated mean and variance of the target variable with two weights for mean and variance that are set by the operator with attention to the importance of each one for a special process. By changing these evidences and calculating the defined score function and maximizing this function with evidences with a normal optimization algorithm, the best values of parent nodes are achieved and the optimization process is done. For increasing the precision of our work, all the nodes of networks are defined as a continuous node but a similar procedure can be done with easier and faster calculations and processing on discrete nodes. With Bayesian Networks there is no need to have formulas or relations between variables. All you have to find is some real or simulation data to build your network with. Another positive point of this work is input variables are used in optimization process within their working range and unlike other optimization tools unreachable or irrational values are not calculated. In our work this algorithm is implemented on two case studies and admissible results are made.