Inner Product Regularization for Stochastic Policies
Thus far, inner product regularization has been explored for deterministic policies that map a state to a single action. However, state-of-the-art imitation learning policies are often stochastic. This section focuses on Gaussian policies, normalizing flows, and diffusion policies. Instead of outputting a single action for a given state, these policies define a state-dependent probability distribution over actions.
Deterministic policies map each state to one action.
Conclusion
Code and videos of the experiments are available on the authors’ website.
Frequently Asked Questions
What does a deterministic policy output? It maps a given state to a single action.
How do stochastic policies represent actions? They define a state-dependent probability distribution over actions.
Which stochastic policy types does the paper focus on? Gaussian policies, normalizing flows, and diffusion policies.
Where are the experiment code and videos available? They are available on the authors’ website.