Inducing desired agent behaviors by selectively sharing information with partially informed agents.
In many settings, it may be possible to enhance performance of an AI system by selectively revealing information to agents that are active in it, either via (direct) communication or (indirect) sensor distribution. To account for such settings, I formulated Information Shaping as the problem of selecting which information to reveal (or conceal) to partially informed agents.
information shaping was first introduced in AAAI’20, where it is used to facilitate the ability to recognize the goals of partially informed agents.
I am currently using information shaping in a collaborative setting, which we call Helpful Information Shaping, where the objective is to use a limited communication channel with a partially informed planning agent to guarantee that it can achieve its goal. To represent a variety of partially informed agents with different performance preferences, our KR’2020 paper explores the trade-off between plan robustness and plan cost and suggests a new planner that generates plans with a specified level of robustness.
From a theoretical point of view, information shaping is challenging because the design space is large, because communication with agents may be limited or noisy, and because information provisioning can have a non-monotonic effect on performance. To efficiently find an optimal design I formulate this as a search problem and develop various admissible heuristics that guarantee an optimal solution is found.
We are also using information shaping as part of a novel approach to task planning for robots. This approach not only offers more efficient plans by changing the way connectivity information about the environment is collected and maintained, but also offers the first way to account for user-defined performance preferences in robotic planning. The benefits of this Task-Aware Waypoint Sampling approach are demonstrated on a set of simulated manufacturing tasks in an automated factory