Agents that Converse

该研究项目旨在弥合特定任务人工智能和会话人工智能模型之间的差距。一方面，特定任务的人工智能模型在物流规划、微芯片布局优化、游戏和各种控制问题等许多领域已经达到了超人的能力，但在很大程度上仍然是黑箱模型，很少有途径获得解释或协同完善提出的解决方案。另一方面，会话人工智能模型通过对话自然地与人类交互，但在特定任务的人工智能擅长的目标导向行为方面，通常缺乏超人的能力。通过这种对话式AI和特定任务AI的结合，我的目标是参与以下一些问题。我们如何使用自然语言向在环境中活动的学习代理发出请求？我们如何训练一个在环境中活动的代理，使其能够用自然语言为自己的行为提供理由或解释？我们如何通过补充领域相关的自然语言数据来提高智能体从经验数据中学习的性能？为了做到这一点，我将探索使会话AI代理更擅长目标导向行为的途径，即使在本质上不是基于自然语言的任务中，也可以通过利用我们已经知道的关于训练高性能任务特定AI的途径。一种方法是通过联合训练一个特定任务的人工智能模型和一个对话模型，使它们对世界的了解是兼容的。其他方法包括从一个强大的预训练会话模型开始，并对其进行微调，使其擅长通常由特定任务的人工智能执行的任务。对于后一种方法，需要进行研究，以便学习如何在保留模型的会话能力的同时做到这一点。因为，一般来说，训练模型学习新技能往往会使他们忘记以前学过的技能。通过训练这些模型，我们将能够获得更具可解释性和协作性的模型。在人工智能模型越来越多地用于做出影响人类生活的决策的时代，这一点变得越来越重要。为了使利用人工智能决策的整体过程更加公平，有必要使这些决策更具可解释性。此外，让这些人工智能进行对话将改变模式，从一个模型做出一个最终无可争议的决定，到一个人们可以与模型协同工作，直到达成最终决定的模式。该项目属于EPSRC人工智能和机器人研究领域，在Jakob Foerster教授的监督下进行

This research project aims to bridge the gap between task-specific AI and conversational AI models. On one hand, task-specific AI models have reached super-human ability in many areas such as logistics planning, microchip layout optimization, games and various control problems, but remain largely black-box models, with little avenue for obtaining explanations or collaboratively refining proposed solutions. On the other hand, conversational AI models naturally interface with humans through conversations, but usually lack superhuman capability in the goal oriented behaviours that task specific AIs excel at.Through this combination of conversational AI and task-specific AI I aim to engage with some of the following questions. How can we use natural language to make requests from learnt agents acting in an environment? How can we train an agent acting in an environment to be able to provide justifications or explanations for their actions in natural language? How can we improve the performance of an agent learning from experiential data by supplementing it with domain-related natural language data?To do this I will explore avenues for making conversational AI agents more adept at goal oriented behaviour, even in tasks that are not natural language-based in nature through avenues that utilise what we already know about training performant task-specific AI. One approach would be through jointly training a task-specific AI model with a conversational model in a way that the representations they learn about the world are compatible. Other approaches involve starting with a powerful pretrained conversational model and fine-tuning it to excel at a task that is usually performed by task-specific AI. For the latter kind of approach, research is needed in order to learn how to do this while retaining the model's conversational abilities. Because, in general, training models to learn new skills tends to make them forget previously learnt skills.Through training these models we will be able to obtain models that are more interpretable and collaborative. This is becoming increasingly more important in an age where AI models are increasingly used in making decisions that affect human lives. Making these decisions more interpretable is necessary for making overall processes that utilise AI-decision-making fairer. Furthermore, making these AIs conversational will change the paradigm from one where a model makes a single final indisputable decision, to one where people can work collaboratively with a model until a final decision is reached.This project falls within the EPSRC AI and Robotics research area and is conducted under the supervision of Professor Jakob Foerster