Otimização de agentes autônomos no AWS DeepRacer: uma abordagem utilizando o algoritmo PPO e trajetórias K1999

Abstract

This work presents the development and validation of a comparative methodology for optimizing autonomous agents on the AWS DeepRacer platform, contrasting Baseline reward approaches with apex-tracking strategies using K1999 trajectories. The objective was to evaluate the impact of offline track preprocessing on PPO algorithm performance and stability. A Streamlit tool calculated ideal racing lines and speed profiles. Empirical data revealed a significant trade-off: while the K1999 trajectory provided the lowest absolute lap time on high-complexity circuits (17.74s on Tokyo Training Track), geometric constraints resulted in dynamic instability. In contrast, Baseline models demonstrated superior consistency and dominated generalization tests on the unseen Summit Speedway circuit. It is concluded that K1999 acts as a performance maximization tool for known circuits, whereas conventional training ensures adaptability.