Optimizing AI Pipelines: A Game-Theoretic Cultural Algorithms Approach

摘要

The structure of contemporary AI applications in complex automation domains, such as robotics and autonomous driving, is multi-staged and hierarchical. The overall pipeline consists of perception, planning, and actuation subsystems. Each of these in turn, consists of staged processing. Such systems consume raw sensor data, and process it to respond intelligently to their surroundings, in the pursuit of assigned goals. Further, such systems use a variety of techniques including signal processing, computer vision, machine learning and `traditional' AI methods (e.g. rules engine, planning, and scheduling, etc.). There may be complex inter-and intra-pipeline interactions that are governed by 100's of tunable parameters, yielding a highly complex system. Optimizing the system-level performance of such complexly interacting subcomponents, is a major challenge for the industry. This paper attempts to address such a challenge with the application of a knowledge-intensive evolutionary optimization framework-Cultural Algorithms. A key component of Cultural Algorithms-which are modeled after problem solving processes in social networks-is the mechanism for distributing knowledge in the population network. Here a new, game-theoretic knowledge distribution mechanism is devised which supports both cooperation and competition between players. The performance of this new mechanism is compared against the de-facto Weighted Majority Win, purely competitive mechanism on a real-world, computer-vision based AI pipeline that supports of autonomous driving. The preliminary results suggest that a game-the-oretic approach is better at combining the workflow stages of the pipeline so as to improve driving behavior than the traditional competition-based approach.