General forward model

研究情境人工智能(AI)算法，如深度学习，已经成功地应用于计算机视觉、自然语言处理和决策系统等许多领域。人工智能尤其成功地学习了如何玩游戏，如中国古代围棋游戏，雅达利游戏，如突破和乒乓球，以及最近的复杂游戏，如星际争霸和DOTA2。许多成功的算法需要有效地预见时间，并根据这些内部实验的结果规划下一步棋的能力。这通常需要访问系统的完美底层模型，或者底层系统的非常精确的手工制作的复制品。最近有几次尝试实际建立模型来预测可能的未来结果，并相应地制定计划，而不需要访问底层模型。目的和目标本研究的目的是创建一种有效和通用的方法来准确预测模拟环境。这项研究将回答几个问题：模拟环境的预测模型能否改善训练结果？训练人工智能算法执行复杂任务是否需要准确的环境模型？或者，是否可以使用模型来有效地过滤不必要的信息？在有几个玩家的游戏中，预测模型是否可以用来创造更好的对手？模拟环境的预测模型是否可以用来理解风险？潜在的应用和好处将游戏作为研究的试验台在安全性、成本和时间控制方面具有优势。例如，在训练自动驾驶汽车避免碰撞时。在模拟环境中，汽车可以在不危及公众的情况下出错；它们可以完全存在于软件中，不需要维护或更换物理组件；模拟可以加快速度，这意味着几个小时的训练时间可以压缩到几分钟。如果有方法可以在现实环境中预测结果，那么这项研究的应用就很广泛。能够准确地预测未来几秒钟后的操作结果将为安全关键应用程序提供巨大的优势。此外，在参数难以预测的环境中，能够提前计划最好和最坏情况的人工智能算法将能够理解其行动中的潜在风险。

Reasearch ContextArtificial intelligence (AI) algorithms such as deep learning has been used successfully in many fields such as Computer Vision, Natural Language Processing and Decision making systems.AI has been particularly successful in learning how to play games such as the ancient chinese game of Go, Atari games such as Breakout and Pong, and more recently complex games such as Starcraft and DOTA2.Many successful algorithms require an ability to effectively see-forward in time, and plan next moves based on the outcome of these internal experiments. This usually requires access to a perfect underlying model of the sytem, or a very accurate hand-crafted replica of the underlying system. There have been several attempts recently to actually build models to predict the possible future outcomes and plan accordingly without access to the underlying model.Aims and ObjectivesThe aim of this research is to create an efficient and genric way for simulated environments to be accurately predicted. There are several questions that will be answered by this research:Can predictive models of the simulated environment improve the results of training?Are accurate models of the environment necessary to train AI algorithms to perform complex tasks? Or can models that effectively "filter" unecessary information be used?In games with several players which amy or may not have predictable actions, can predictive models be used to create better adversaries?Can predictive models of simulated environments be used to understand risk?Potential Applications and benefitsUsing games as a test-bed for research has advantages in safety, cost and control of time. For example, when training self-driving cars to avoid collisions. In a simulated environment,the cars can make mistakes without endangering the public; they can exist entirely in software and no physical components need to be maintained or replaced; simulations can be 'sped up' meaning that several hours of training time can be compressed into minutes.If there are ways to predict outcomes in real-world environments, then the applications of this research are widespread. Being able to accurately predict the outcomes of actions several seconds into the future would provide a large advantage for safetly-critical applications. Additionally in the context of environments with hard to predict parameters, AI algorithms that can plan ahead for best and worse case scenarios would be able to understand underlying risk in it's actions.