Task assignment strategy has a great impact on the performance of the workflow management system. The instability of human resource brings challenges to task assignment. General task assignment strategies have some deficiencies. First, they only consider the individual attributes of candidate resources, ignoring the influences to the candidate resources from other resources in process. In addition, they need to setup a capability index of each resource in advance. However, it is hard to make the capability index fit the actual situation, and a wrong capability index will make the workflow engine assign the task to the unsuitable resource, degrading the performance of workflow management system. To overcome the above deficiencies, four Q-learning-based task assignment algorithms are proposed according to different state transition views and different reward functions. Simulation experiments show that Q-learning-based task assignment algorithms can work well even without setting up a capability index in advance. Also due to their support to consider the social relationship, the average time of case completion decreases.