Reproduce “Meta-Cognition. An Inverse-Inverse Reinforcement Learning Approach for Cognitive Radars” on OpenResearch

About this paper

Kunal Pattanayak, Vikram Krishnamurthy, Christopher Berry

Cornell University researchers developed an Inverse-Inverse Reinforcement Learning (I-IRL) framework for cognitive radars, enabling them to strategically obfuscate their utility functions from intelligent adversaries. This approach deliberately introduces sub-optimal behavior to confuse adversarial Inverse Reinforcement Learning (IRL) algorithms, achieving significant increases in adversary confusion with relatively small radar performance losses.

Key insights

• Meta-cognition involves a radar being aware of an adversary's IRL efforts and deliberately modifying its behavior to obfuscate its true objectives.
• The I-IRL framework extends Afriat's theorem to noisy observations, enabling adversaries to perform statistical hypothesis tests for utility maximization with quantifiable Type-I error probabilities.
• Strategic sub-optimality, or a deliberate performance loss, can effectively prevent adversaries from accurately detecting or reconstructing the radar's underlying utility function.

Summary by alphaXiv.