Provided is a highly reliable technology for optimizing an action for controlling an environment in a target space. An action optimization device for optimizing an action for controlling an environment: acquires environmental data related to a state of the environment; performs time/space interpolation on the acquired environmental data; trains an environment reproduction model, based on the time/space-interpolated environmental data, such that, when a state of an environment and an action for controlling the environment are input, a correct answer value for an environmental state after the action is output; trains an exploration model such that an action to be taken next is output when an environmental state output from the environment reproduction model is input; predicts a second environmental state corresponding to a first environmental state and a first action by using the trained environment reproduction model; explores for a second action to be taken for the second environmental state; and outputs a result of the exploration.