具有自我調適特性的電腦輔助群眾計算之研究

Abstract

利用人力介入計算過程以解決電腦難以提供滿意結果的問題是近年來新興的研究領域之一。人類本身具有的知識與推理歸納能力可以大幅降低設計解決方案的成本。然而，人工計算過程無法擁有與電腦計算時相同的長期穩定性，特別是當此人力資源是透過網際網路上專業技能差異甚大的群眾所提供。因此，探討人力介入計算過程中對於整體表現產生的影響，與如何更進一步在考慮人與機器個別優缺點的情況下，提供人機合作框架以追求高質量的結果是很重要的議題。在本論文中，我們首先探討在人智計算系統 (human computation system)中的對於效能表現的主要影響因子，然後針對一般化計算問題並同時考慮人機合作情況下，定義一個效益最佳化問題。基於所提出的效益最佳化問題，人智計算系統中的利弊權衡 (tradeoff)也呈現在此篇論文中。為了支持人機合作，我們先定義了人機可交換資訊的形式 (human/machine exchangeable information)以描述人的解題行為與工作目標之間的關係，並藉此提出在動態環境中具有自我調適特性的電腦輔助群眾計算的框架。更進一步針對三種具代表性的人智計算系統：集體智慧 (collective intelligence)、社交計算 (social computing)與群眾外包 (crowdsourcing)，提出相對應的解決方案與人機可交換資訊模型。動態工作流程模型用以描述使用者即時行為與測試目標之間的關聯性；基於易於解釋的屬性之使用者特徵模型可以幫助擷取出精確的社交資訊；以教學為目的之反釣魚攻擊知識本體論可以用更簡單的方法來建立。實驗結果顯示，我們所提出的模型框架可以有效的利用群眾偏差，並藉此提升群眾計算的表現。

Human computation has attracted considerable attention in reducing design effort of solving problems which cannot be fully characterized by limited attributes. However, the involvement of human solvers in computation process leads to unstable performance, especially when solvers are recruited via Internet. In this dissertation, we first explore factors affecting the performance of human computation system and then define the utility optimization problem of a generalized computational system considering both automatic and manual problem-solvers. Based on the problem formulation we proposed, we analyze the tradeoffs in the context of three types of human computation systems: collective intelligence, social computing and crowdsourcing. Afterwards, we define the form of Human/Machine Exchangeable Information describing the relationships between users’ runtime behaviors and task objectives, and then present self-organizing computer-aided frameworks of crowd computation in a dynamic environment accordingly. Therefore, precise social information can be extracted via user characteristics model from the easy-to-explain feature space and anti-phishing attack knowledge ontology can be easily built for pedagogical purpose. The experimental results show that our proposed frameworks can utilize crowd’s bias effectively and hence improve the performance of crowd computation.