Adaptive Risk-Sensitive Optimal Control with Dual Features Through Active Inference

We propose a tractable adaptive risk-sensitive optimal control framework tailored to uncertain nonlinear stochastic system dynamics. The architecture exhibits dual features meaning that the controller actively maintains a balance between exploitation and exploration. The problem statement is cast as an instance of Active Inference. Active Inference is an emerging framework in theoretical neuroscience that seeks to explain the behaviour of biological agents by practising inference on probabilistic graph models. The developed algorithm leverages a receding horizon strategy that simultaneously estimates the uncertain parameters of the dynamic system from past observations and designs controller parameters by predicting the future performance of the controlled system. The algorithm does not make use of the separation and certainty equivalence principles. We further show that for the special case of linearly parameterized controller and dynamics, the approach leads to a quadratic programming problem maintaining a manageable computational complexity. The capability and anticipated properties of the proposed algorithm are demonstrated on a simulated nonlinear system.