A multilayer affective computing model with evolutionary strategies reflecting decision-makers’ preferences in process control

Many industrial control problems related to multi-objective optimization, such as controller parameters tuning, often require operators to perform multiple-step interactions without considering the changes of decision-makers’ affective states and quantitative description of decision-makers’ preferences during the interactive decision. Regarding this problem, we developed a multilayer affective computing model (MACM), including three factors: human personality, emotional space, and affective states, to demonstrate the iterative affective computing during the interactions. First, a concise model of affective computing-driven interactive decision-making was built before three submodules involved were described in detail. (1) An affective state recognition method based on facial expressions was presented, providing the basis for obtaining expert affective states during decision-making. (2) An identification method of affective parameters was given, providing an approach to describing personalized affective state-changing rules of different persons. (3) A definition of decision-makers’ preferences in interactive decision-making was specified. In addition, a decision-makers’ preferences mining method was developed by the MACM and an iterative learning control (ILC) strategy. Thus, we proposed affective computing-driven interactive decision-making method, which provided a simplified approach to converting the interactive decision problems based on decision-makers’ preferences to decision issues based on incremental decision vector, along with assisting computers to learn from human experts and perform decision-making automatically in a general sense. Then, two typical process control cases—PI controller tuning for decoupling problem and manipulate vector optimization for batch processes were used to show the correctness and effectiveness of the contributions. Compared with other traditional optimization algorithms without affective state tracking and recognition (fuzzy control, ILC, reinforcement learning, and so on), experimental results indicated that the proposed method could achieve good performance. Finally, this study presented the efficiency and limitations of using this technique for a specific application