A learning method for learning an action of an agent using model-based reinforcement learning is provided. The learning method includes: obtaining time series data indicating states and actions of the agent when the agent performs a series of actions; establishing a dynamics model by performing supervised learning using the time series data obtained; deriving a plurality of candidates for an action sequence of the agent from variational inference using a mixture model as a variational distribution, based on the dynamics model; and outputting, as the action sequence of the agent, one candidate selected from among the plurality of candidates derived.