Abstract: In view of existing ARAS (adaptive response by attractor selection) models’ shortcomings that was noise item produces an effect ambiguously and activity couldn’t determine the opportunity of state changing correctly, an improved ARAS model was proposed. The model specifies the stochastic effect of the noise on the system by an attractor with a noise item. Through putting forward an activity formula inversely proportional to the partial difference between current path status and best path status, it was validated by Matlab with the improved model could increase the accuracy of load balancing in a simple case. A multi-path routing by ARAS mechanism (MPAM) based on AODV is proposed, which selects routing path with the improved ARAS model. Furthermore, the improved model is validated to be able to reduce the average network delay by OPNET.