Model-Free Inverse H-Infinity Control for Imitation Learning

This paper proposes a data-driven model-free inverse reinforcement learning (IRL) algorithm tailored for solving an inverse
$H_\infty$
control problem. In the problem, both an expert and a learner engage in
$H_\infty$
control to reject disturbances and the learner’s objective is to imitate the expert’s behavior by reconstructing the expert’s performance function through IRL techniques. Introducing zero-sum game principles, we first formulate a model-based single-loop IRL policy iteration algorithm that includes three key steps: updating the policy, action, and performance function using a new correction formula and the standard inverse optimal control principles. Building upon the model-based approach, we propose a model-free single-loop off-policy IRL algorithm that eliminates the need for initial stabilizing policies and prior knowledge of the dynamics of expert and learner. Also, we provide rigorous proof of convergence, stability, and Nash optimality to guarantee the effectiveness and reliability of the proposed algorithms. Furthermore, we showcase the efficiency of our algorithm through simulations and experiments, highlighting its advantages compared to the existing methods.
Note to Practitioners
—Generally, the cost function for optimal tracking or imitation control is manually defined, which is a challenging task and may result in large tracking errors and slow tracking. In such cases, IRL is a powerful tool for reconstructing proper cost functions. Real-world systems, as demonstrated in practical cases, are frequently exposed to external disturbances and come with unknown models. Employing
$H_\infty$
control is an effective strategy to handle disturbances. However, applying model-free IRL to solve the inverse problem of
$H_\infty$
control for imitation remains an underexplored domain. This paper explores model-free inverse
$H_\infty$
control for imitating expert behaviors, specifically addressing the time-consuming nature of the existing IRL studies that employ a two-loop iteration structure. We propose an efficient single-loop IRL algorithm with a new framework to do this. It is data-driven and model-free, eliminating the need to find an initial stabilizing control policy, which is typically challenging. Additionally, it ensures convergence, stability, and optimality with provable guarantees.