Expected free energy

Expected free energy is the quantity that an active inference agent minimizes when selecting actions. It combines two costs: the pragmatic cost of failing to achieve preferred outcomes, and the epistemic cost of failing to resolve uncertainty about the hidden states of the environment. The expected free energy of a policy is not the same as the free energy of the present moment; it is a forecast, a prediction about how much free energy the agent will have in the future if it follows that policy. In this sense, expected free energy is the bridge between the statics of inference and the dynamics of action. The agent does not merely minimize surprise about the present; it minimizes expected surprise about the future. The concept is mathematically elegant but computationally demanding: to compute expected free energy for even a simple policy requires integrating over all possible future trajectories. Real brains almost certainly do not compute expected free energy exactly. They approximate it, and the quality of behavior is the quality of the approximate inference.