Developmental reinforcement learning

发展强化学习

• 批准号: 238988-2010
• 负责人: Precup, Doina
• 金额: $ 4.37万
• 依托单位: McGill University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2016
• 资助国家: 加拿大
• 起止时间: 2016-01-01 至 2017-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=594210
• 关键词: Developmental; reinforcement; learning

We propose to address the major problem of building artificial intelligent agents that exist for an extended period of time, during which they continually improve their world representations. We will use the theoretical framework of reinforcement learning, which allows agents to learn by interacting with an unknown, stochastic environment, and receive reward feedback form it. Unlike in existing reinforcement learning methods, we will not try to infer a true environment state. Instead, agents will build their own internal state, which will be sufficient to make predictions. Prediction errors will be used to guide the agents' exploration, in a process mimicking curiosity.

我们建议解决的主要问题，建立人工智能代理存在的一段时间，在此期间，他们不断提高他们的世界表示。我们将使用强化学习的理论框架，它允许智能体通过与未知的随机环境交互来学习，并从中获得奖励反馈。与现有的强化学习方法不同，我们不会尝试推断真实的环境状态。相反，智能体将建立自己的内部状态，这将足以进行预测。预测误差将被用来指导智能体的探索，这是一个模仿好奇心的过程。