Scholarlink

In the rapidly evolving domain of reinforcement learning (RL), which has applications in computer vision and games, our research presents an RL-based Embedding algorithm (EmbRL) that, applied to an autonomous car racing environment, allows for rapid algorithm training with significant results. EmbRL addresses the challenge of processing high-dimensional camera inputs, which is common in advanced game environments like OpenAI Five and AlphaStar. By employing a pre-trained supervised learning model, our algorithm efficiently transforms these inputs into a set of 1000 class features, which are then processed by a fully connected network (FCN) acting as the RL model. This method effectively separates the task of understanding the vehicle’s state from the core path-finding and control tasks performed using a separate RL network, simplifying the task of autonomous car racing. Our findings show a remarkable reduction in training time, speeding up training by 230% compared to traditional end-to-end convolutional networks, as well as a significant boost to the reward, highlighting EmbRL’s potential in enhancing the real-time applicability of vision models. This study integrates concepts from established methodologies, incorporating minor modifications that result in significant enhancements in performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Autonomous Vehicle Racing Using Reinforcement Learning with Pre-trained Embeddings for Dimensionality Reduction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Martin Holen,
  • Jayant Singh,
  • Christian W. Omlin,
  • Jing Zhou,
  • Kristian M. Knausgård,
  • Morten Goodwin

摘要

In the rapidly evolving domain of reinforcement learning (RL), which has applications in computer vision and games, our research presents an RL-based Embedding algorithm (EmbRL) that, applied to an autonomous car racing environment, allows for rapid algorithm training with significant results. EmbRL addresses the challenge of processing high-dimensional camera inputs, which is common in advanced game environments like OpenAI Five and AlphaStar. By employing a pre-trained supervised learning model, our algorithm efficiently transforms these inputs into a set of 1000 class features, which are then processed by a fully connected network (FCN) acting as the RL model. This method effectively separates the task of understanding the vehicle’s state from the core path-finding and control tasks performed using a separate RL network, simplifying the task of autonomous car racing. Our findings show a remarkable reduction in training time, speeding up training by 230% compared to traditional end-to-end convolutional networks, as well as a significant boost to the reward, highlighting EmbRL’s potential in enhancing the real-time applicability of vision models. This study integrates concepts from established methodologies, incorporating minor modifications that result in significant enhancements in performance.