將機器心智理論模型應用於複雜社交網路

Abstract

社交網路定義了人類的人際關係。然而人際關係是抽象且需要觀察才能了解的，因此若要描繪社交網路的結構，需要藉由觀察社交行為方能建構。建構社交網路的結構對人類而言十分自然，我們汲取社交行為中的資訊並藉此平順地與人互動。無法與人類自然的產生社交行為成為發展流暢的人機互動的一大挑戰。在這篇研究中，我們認為要讓機器擁有社交能力的必要條件是，機器必須有能力藉由觀察社交行為推測社交網路的結構。在先前的研究中，我們設計一機器心智理論模型（此機器心智理論模型的雛型來自於 Google Deepmind），此模型可藉由觀察代理人（agent）的社交行為推測出該代理人對環境中其他四個人的友好程度。然而在更加複雜的社交網路中，此模型的表現仍然未知。在觀察富有動態變化的社交行為時，模型需要整合更多層面的資訊去推測社交網路的結構。在這篇研究中，我們擴展了機器心智理論模型，擴展後此模型具備能夠解讀更動態的五位代理人組成的社交行為。我們設定ㄧ組社交網路並定義移動規則，使五位代理人遵照移動規則於24x24的網格世界中自由移動，並將移動之最後10個步驟（包含最終位置）記錄起來，用於訓練此機器心智理論模型。擴展模型的過程中，為了符合訓練資料的需求，我們將模型原先的張量（tensor）從12*12*11*10擴大為12*12*31*10，深度增加至31是為了正確傳遞代理人在網格世界中的移動資訊。我們使用了具有五位代理人但未被訓練的網格世界做為測試模型學習的結果，並使學習後的模型輸出對五位代理人最終位置的預測。為了評估模型學習的效果，我們註記每一組預測結果中，具有相鄰代理人的情形，並疊加了所有1000個測試網格世界的預測結果。另外，我們將此模型同步測試於已學習社交網路的測試網格世界資料，以及另外100組未被此模型學習的社交網路的網格世界資料，此模型可以正確解讀與分辨所學習的社交網路與另外100組社交網路的不同。這些結果顯示，此機器心智理論模型得以解讀抽象而且更加複雜的動態社交網路關係，並有潛力更廣泛應用於社交機器人（socially assistive robots）。

Social networks define the structure of human inter-relationships. Critically, such structures are abstract and hidden such that to learn about them requires the ability to infer hidden states from observations of social preference behaviors. Human beings do this very naturally in using inferred social knowledge for smooth daily communication and interaction with each other. A key challenge in human robot interaction (HRI) is to imbue machines with similar ability to engage with humans naturally. In this study, we consider that in order to design machines that interact effectively with humans, the artificial intelligence driving these machines should also be able to infer social networks from observations of human social behaviors. We previously designed a Machine Theory of Mind (ToMNet+; modified based on principles applied in Google DeepMind projects) that displayed the ability to observe social interaction behaviors and infer simple human social networks consisting of one agent with distinct socializing preferences for four targets. However, how this model performs for more complex social structures remains unclear. Critically, more dynamic social structures driving social behaviors requires the model to incorporate more dimensions of information towards inferring social networks. In this present study, we expanded ToMNet’s structure and built ToMNet 2.0. ToMNet 2.0 has the capability to infer more complex human social networks consisting of five dynamically interacting agents each with its own set of social preferences for the other four agents in the social network. Such a social preference setting drives all five agents to move dynamically in 24 x 24 grid world over 10 time steps based on a movement formula, which is then utilized as ToMNet 2.0’s training data. To meet the social interaction information dimensionality of this dynamic grid world data, we adjusted ToMNet 2.0 to accept input tensor sizes from 12*12*11*10 (in the original model, 11 channels coded the 5 agents positions and barriers) to 12*12*31*10, the increased depth of the tensor size to 31 conveys the additional information needed to code the dynamic moving agents. At test, ToMNet 2.0 evaluates the start states of agent locations in novel test grid worlds and predicts the final locations of the five agents relative to each other. To evaluate ToMNet 2.0’s performance, we summed the number of times agents clustered with others at the final states in 1000 start state permutations. We compared ToMNet 2.0’s performances for trained social networks and 100 untrained permuted social networks against the ground truth final state determined by the movement formula (details provided in methodology). These results show that ToMNet 2.0 is able to dissociate the hidden social preferences of the five agents from the 100 untrained social networks. Thus, ToMNet 2.0 is able to utilize additional social preference information compared with ToMNet, which raises the potential for ToMNet 2.0 to be applied in socially assistive robots problems more broadly.