In the most computer networks researches, the algorithm performance is depends on the value of the thresholds. The mathematical relations are the best and most accurate method to calculate the thresholds. While the most issues are complex and heuristic, modelling with mathematical relations is very difficult or impossible. In this way, another method is trial and error that the values are generated by simulation and the best ones are selected. These values are appropriate for the current state of the network and the values must be recalculated if the conditions change. When the number of parameters is increased, this method is more time consuming. In such circumstances, the learning method could be used as a good alternative mechanism. In this paper, instead of traditional methods to calculate the thresholds, the learning approach is selected to obtain the threshold values to control the Session Initiation Protocol (SIP) overload. This case study emphasizes the ability of the new method. SIP is a standard protocol to use in Next Generation Network (NGN) architectures but it does not have suitable mechanisms to handle overload. The basic mechanism embedded in SIP cannot eliminate the problem by rejecting requests, which cause collapse in network. This challenge creates a sharp drop in quality of service for NGN users. Therefore, many algorithms have been proposed to overcome overload in SIP network, among which, multi-agent systems are the new approach. Multi-agent system is a powerful tool to model and develop the complex large scale distributed systems. Since the SIP networks have the same properties, multi-agent is a good selection for SIP overload control. The due complexity can be reduced by holonic organization. In proposed method, the threshold values are generated by Q-learning. The simulation results demonstrate that the Q-learning output overcomes the previous method. The method also increases total throughput, reduces delay, and considers fairness in the S
