Scholarlink

In this work, we designed a target selection test in order to investigate the most appropriate metric to measure the quality of an immersive experience. The test was designed in Unity, and instead of usual subjective tests with a group of volunteers, we incorporate to it a reinforcement learning system which represents an impartial agent to carry out the tests. The idea is that the network will learn to hit the target along the training and grows its precision when evaluated by a quality metric. The precision will increase until it reaches its maximum score. If we define different metrics, we can analyze the behavior from the beginning of the training until it reaches its maximum value, and then we define the best metric. We use three quality equations that provide a performance score ranging from 0 to 1 with the following behavior when they vary from along the range: (1) a linear variation, (2) a positive concavity exponential, and (3) a negative concavity exponential variation. Experimental results show that metric 1 presents the best results, followed by metrics 2 and 3 respectively.

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

Automated Evaluation of Metrics for Immersive Test Using Reinforcement Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Roberto G. Estrada L.,
  • Anderson V. C. de Oliveira,
  • Agustin Alejandro Ortiz Diaz,
  • Jeferson B. da Costa,
  • Emerson S. Domingos

摘要

In this work, we designed a target selection test in order to investigate the most appropriate metric to measure the quality of an immersive experience. The test was designed in Unity, and instead of usual subjective tests with a group of volunteers, we incorporate to it a reinforcement learning system which represents an impartial agent to carry out the tests. The idea is that the network will learn to hit the target along the training and grows its precision when evaluated by a quality metric. The precision will increase until it reaches its maximum score. If we define different metrics, we can analyze the behavior from the beginning of the training until it reaches its maximum value, and then we define the best metric. We use three quality equations that provide a performance score ranging from 0 to 1 with the following behavior when they vary from along the range: (1) a linear variation, (2) a positive concavity exponential, and (3) a negative concavity exponential variation. Experimental results show that metric 1 presents the best results, followed by metrics 2 and 3 respectively.