The paper considers supervised learning algorithm of nonlinear perceptron with dynamic targets adjustment which assists in faster learning and cognition. A difference between targets of the perceptron corresponding to objects of the first and second categories is associated with stimulation strength. A feedback chain that controls the difference between targets is interpreted as synthetic emotions. In a population of artificial agents that ought to learn similar pattern classification tasks, presence of the emotions helps a larger fraction of the agents to survive. We found that optimal level of synthetic emotions depends on difficulty of the pattern recognition task and requirements to learning quality and confirm Yerkes-Dodson law found in psychology.