Cumulative Contradictory Expected Regret

The primary idea behind regret-based robustness is to minimize the variance in outcomes for an agent while retaining good nominal performance. Our solution, the family of Robust Adversarial Defense (RAD) algorithms, optimizes a novel notion of regret that is scalable and principally-motivated.

Github LinkedIn G.Scholar

CCER

Computing the expected regret over entire trajectories is an intractable problem, so we make two key improvements in our method. First, we restrict the set of states covered in the regret expectation to an \(\epsilon\)-bounded neighborhood, which is a common concession in adversarial attacks and robustness (the rationale being, a sufficiently large perturbation would be detectable and thus ignored). Secondly, we provide scalability through myopic expectations of regret, i.e. at each state we compute the regret over potential mistakes only at this time step, and accumulate them in a Q-learning sum. Mathematically, this retains the markov assumption and reduces the computational complexity from combinatorial O(n!) to a constant scalar O(kn). Empirically, this "smooths" out agent behavior as the model seeks trajectories with a margin for error.

Formally, we optimize the difference in rewards for the nominal (unperturbed) observation \(s\) and the expected adversarial neighborhood \(N(s)\).

\[\delta(s,a)=R(s,a)-\sum_{\hat{s} \in N(s)}R(\hat{s},a) + \gamma \mathbb{E}_{s^{'}}\delta(s^{'},\pi(s^{'}))\]

Optimizing CCER with RAD

We use DQN and PPO algorithms to optimize the CCER metric, yielding robust RL models. In our experiments, we noticed that the robust models, while stable, tended towards ultra-conservatism in their actions. For example, a vehichular agent in a busy intersection setting (highway-env/intersection-v0) would opt never to move instead of risking an unprotected turn. We addressed this by optimizing both CCER and reward with a 1:4 weighting, which we found to be near optimal across all environments.

Get in touch

If you want to collaborate, have a question, or just want to say hi, you can contact me through the links below.