Inference is one of human's high-level functionalities and it is not easy to implement in machine. It is believed that inference is not results of single neuron's activity. Instead, it is a complex activity generated by multiple neural networks. Unlike computer, it is more flexible and concludes differently even for the similar situations in case of human. In this paper, these characteristics are defined as "informality." Informality in inference can be implemented using the interaction of multiple neural networks with the inclusion of internal or subjective properties. Simple inference tasks such as pattern recognition and robot control are solved based on the informal inference ideas. Especially, fuzzy integral and behavior network methods are adopted to realize that. Experimental results show that the informal inference can perform better with more flexibility compared to the previous static approaches.