Our brain is a complex and not yet fully understood system. Trying to figure out how our mind functions is something that has been pursued by philosophers for as long as written evidence can prove. In this context, some of the most intriguing questions that can arise are related to the possibility to find out what is happening inside someone else’s (or even one’s self) mind. Is human behavior predictable given enough variables are observable?

我们的大脑是一个复杂且尚未完全了解的系统。 只要书面证据能够证明，试图弄清我们的思维方式是哲学家一直追求的东西。 在这种情况下，可能会出现的一些最引人入胜的问题与发现别人(甚至一个人的自我)内心正在发生的事情的可能性有关。 如果可以观察到足够多的变量，人类行为是否可以预测？

Such reflections stimulated psychologists to develop what is so-called “Theory of Mind” (ToM), as described by AI Goldman [1]:

这些反射刺激了心理学家发展出所谓的“心灵理论”(ToM)，正如AI Goldman [1]所述：

‘Theory of Mind’ refers to the cognitive capacity to attribute mental states to self and others. Other names for the same capacity include “commonsense psychology,” “naïve psychology,” “folk psychology,” “mindreading” and “mentalizing.” […] How do they [people], or their cognitive systems, go about the task of forming beliefs or judgments about others’ mental states, states that aren’t directly observable?

“心理理论”是指将心理状态归因于自我和他人的认知能力。 具有相同能力的其他名称包括“常识心理学”，“幼稚心理学”，“民间心理学”，“心态”和“心理化”。 […]他们[人]或他们的认知系统如何去完成对他人的心理状态，不能直接观察到的状态的信念或判断的任务？

Different empirical tests were proposed (e.g., [2], [3]) to examine if observable inputs can lead to a mental state prediction and if knowing the mental state allows to predict future actions (behavior). While psychologists still debate if understanding the state of mind is even possible for us humans, ToM has attracted Artificial Intelligence researchers’ attention.

提出了不同的经验检验(例如，[2]，[3])以检查可观察的输入是否可以导致心理状态预测，以及了解心理状态是否可以预测未来的行为(行为)。 尽管心理学家仍在争论是否甚至对我们人类而言都可能了解心理状态，但ToM吸引了人工智能研究人员的注意。

心智与人工智能理论 (Theory of Mind and Artificial Intelligence)

AI has evolved rapidly in the last decade taking advantage of the increased computation power and, particularly, with the breakthroughs achieved with deep neural models. Although powerful, such results are achievable in specific settings and the obtained models can perform well only in the tasks they were designed to excel at. The next step would be to design an AI able to deal with a whole range of different problems and learn while interacting with different entities, as highlighted by 2 of the 10 grand open challenges in the field of robotics [4]:

在过去的十年中，人工智能已经得到了Swift发展，它利用了不断增强的计算能力，尤其是深层神经模型所取得的突破。 尽管功能强大，但在特定的设置下仍可获得这种结果，并且所获得的模型仅在他们擅长的任务中才能表现良好。 下一步将是设计一种能够处理各种不同问题并在与不同实体进行交互的同时学习的AI，这在机器人技术领域的10大挑战中有2个尤为突出[4]：

(vi) Fundamental aspects of artificial intelligence (AI) for robotics, including learning how to learn, combining advanced pattern recognition and model-based reasoning, and developing intelligence with common sense.

(vi)机器人技术人工智能的基本方面 ，包括学习如何学习，将高级模式识别与基于模型的推理相结合以及发展具有常识的智能。

(viii) Social interaction that understands human social dynamics and moral norms and that can be truly integrated with our social life showing empathy and natural social behaviors.

(viii)能够理解人类社会动态和道德规范的社会互动 ，并且可以与我们表现出同理心和自然的社会行为的社会生活真正融合。

Combining the concepts and ideas of the ToM with synthetic intelligent systems has proven a viable option to tackle such points and becomes especially relevant when considering multi-agent AI systems and problems with machine-human interaction. Different works have been built on top of this premise, e.g., using a Bayesian ToM framework in [5] or even inverse Reinforcement Learning to ‘reverse engineer’ an observable behavior, as done in [6]. Below I comment on a different approach proposed by researchers from DeepMind and Google Brain.

实践证明，将ToM的概念和思想与合成智能系统相结合是解决此类问题的可行选择，并且在考虑多主体AI系统和人机交互问题时尤其重要。 在此前提之上已经建立了不同的工作，例如，使用[5]中的贝叶斯ToM框架，甚至使用反向强化学习来“逆向工程化”可观察到的行为，如[6]中所做的。 下面，我评论DeepMind和Google Brain研究人员提出的另一种方法。

互联网 (ToMnet)

In [7], the authors describe what was designated as the Machine Theory of Mind. The idea is to treat ToM as a meta-learning problem (i.e., learn how to learn), to achieve an artificial system with prior knowledge about how an agent it has never met before is going to behave and to collect data about how this agent interacts with the environment to create individual models to encompass particularities in the behavior of specific agents.

在[7]中，作者描述了被称为机器心理理论的东西。 这个想法是将ToM视为元学习问题(即学习如何学习)，以事先了解其从未遇到过的代理如何行为的方式来获得人工系统，并收集有关该代理如何运行的数据与环境交互以创建单个模型，以包含特定代理行为的特殊性。

Different from other approaches, the goal is to learn autonomously using limited data rather than asserting a generative model, observing the agent’s behavior, and reverse engineering its algorithm. The experiments consisted of ToMnet interacting with agents that follow different policies in a gridworld, where a discrete set of possible actions is possible at each step. Firstly, with agents following a random policy, then agents trained with full observability over the environment and finally agents trained with different kinds of partial observability.

与其他方法不同，目标是使用有限的数据自主学习，而不是主张生成模型，观察代理的行为并对其算法进行反向工程。 实验由ToMnet与遵循网格世界中不同策略的代理进行交互组成，在网格世界中，每个步骤都可能有一组离散的可能动作。 首先，对于具有遵循随机策略的代理，然后对代理进行环境完全可观察性训练，最后对代理进行不同种类的部分可观察性训练。

The results show that ToMnet fulfills its goal: to learn a general model that applies to the average behavior and specific models when observing new agents. Unfortunately, a rather large communication channel between the agents was necessary to achieve such results in more challenging situations. With future improvements in both the model and the methods, it is expected that equally impressive results will be obtained while dealing with more complex environments and without direct communication among the agents.

结果表明，ToMnet实现了其目标：学习适用于平均行为的通用模型和观察新代理时的特定模型。 不幸的是，在更具挑战性的情况下，代理商之间必须有相当大的通信渠道才能获得此类结果。 随着模型和方法的未来改进，期望在处理更复杂的环境且无需代理之间直接通信的情况下获得同样令人印象深刻的结果。

The possibilities outlined by this work and other authors combining the Theory of Mind and AI are fascinating. Been able to understand the behavior of complex artificial systems may be an important step to comprehend our minds after all.

这项工作和其他结合了心智理论和人工智能的作者所概述的可能性令人着迷。 毕竟，能够理解复杂的人造系统的行为可能是理解我们思想的重要一步。