A hierarchical reinforcement learning system for automatic disease prediction, and method thereof, including an agent simulator module for simulating the acts of doctors; a user simulator module for simulating the acts of patients; a disease classifier module, and an internal critic module, wherein the agent simulator module includes a master module which is in a high level, and a plurality of worker modules which are in a low level; the plurality of worker modules each acts as a doctor from a specific department, while the master module appoints the plurality of worker modules to interact with the user simulator module for collecting information; wherein the master module activates the disease classifier module to output a prediction result when information collected from the plurality of worker modules is sufficient; wherein the internal critic module is configured for generating intrinsic reward to the plurality of worker modules, judging the termination condition for the plurality of worker modules, and wherein the user simulator module is configured for returning extrinsic reward to the mater module.