| Abstract: | 代理人基方法（Agent-based Methodology﹐ABM），已逐漸成為研究諸如財務市場、公司動態、社會與經濟網絡、都市系統、以及交通動態等社會與經濟複雜適應性系統跨領域議題的工具。本研究計劃旨在對代理人基計算經濟學(agent-based computational economics)之現行發展，做出基礎面及應用面之貢獻。基礎面向之課題，在本研究計劃中以「代理人工程」(agent engineering)統稱之；應用面向之課題，即本計劃中所稱之「市場設計」(market design)。然而，此二課題並非獨立無關，本計劃將依據計算力對等(computational equivalence)法則，完成一個整合軟體代理人(software agents)與真人(human agents)的市場平台，做為一個能夠體現並處理市場設計複雜性的工具。代理人基計算經濟學發展至今日，最有政策意涵的一步，就是在市場設計上之進展，尤其是在電力市場這幾年的研究上，從建模到政策評估，更令人產生了正面的期待。然而同時，文獻上也注意到兩個必須妥善面對的層面，即設計的真實性(validity)與設計的頑韌性(robustness) 。而這兩個層面的問題都和「當事人」(stakeholders)有關，「當事人」是如何地學習、謀略與行動將影響到市場的運作。這點不僅呈現在代理人基電力市場近年來之研究中，也普遍存在於一般代理人基計算經濟學的文獻中。簡而言之，代理人工程有其攸關之重要性 (agent engineering matters) 。而近十年來，人工智慧最活躍的一個支派—計算智慧（Computational Intelligence﹐CI），已被應用在自發性代理人的設計之上，尤其是代理人調適機制這部份。雖然許多計算智慧方法都可用來設計代理人，例如強化學習法（reinforcement learning）以及遺傳演算法（genetic algorithms），但如何選擇適當的調適機制，仍是代理人基方法現階段所必須解決的一個課題，特別是當研究結果可因植入的適應機制而有天壤之別的差異時，這個問題就顯得更為重要。盧卡斯（Robert Lucas）曾建議使用真人所進行的實驗室結果—即所謂的實驗經濟學 (experimental economics)—來幫助解決代理人工程中調適機制的選擇問題(selection of adaptation schemes) ，但是對於甚麼樣的實驗室卻沒有交待，而賽門(Herbert Simon)的研究卻告訴我們實驗室中所能提供的「決策支持」(decision supports)本身就可能影響到決策者的學習行為。因此，本研究將盧卡斯與賽門的論述結合，提出一個可以做為代理人工程的新原則，即「計算力對等」(computational equivalence)原則，簡稱CE。並著手建立符合此一原則的實驗室，稱為CE Lab。CE Lab 將使軟體代理人(software agents)與真人(human agents)可以同時存在於一市場平台中，進而整合目前看似相關，卻又獨立進行的兩個領域：代理人基計算經濟學與實驗經濟學。這不僅使得代理人工程的研究與當前e化的趨勢結合，也使得由此而發展的市場設計能不致與日漸普遍的自動化市場(automated markets)概念脫鉤。 CE Lab的建構基礎有二，皆延續本人研究團隊過去所累積之成果。其一是代理人基計算經濟學的模型及軟體，特別是AIE-DA與AIE-AFM，其二則是與王孫崇博士(本計劃共同主持人) 共同進行的以網路為架構的線上實驗市場—臺灣政治期貨市場(Taipex) 。在第一年的研究中，我們將就現有的規模先做擴充。首先擴充AIE-DA與AIE-AFM中的交易機制及軟體代理人的行為模組，而後逐次在Taipex中加入軟體代理人，並計畫在即將到來的北高兩市市長或立法委員選舉中做初步測試。在第二年的研究中，則將逐漸深化並完成AIE-DA與Taipex兩個市場平台的接軌，建構CE-Lab的雛型。而在第三年的研究中，就市場設計文獻中重要且規模適宜的課題，正式進行CE-Lab的運作。在這三年研究中，也將不時與日本東京理工大學Hiroshi Deguchi教授領導的U-MART團隊互相切磋，學習經驗。
Agent-based Methodology (ABM) has become indispensable to the inter-disciplinary study of social and economic complex adaptive systems, such as financial markets, companies dynamics, social and economic networks, urban systems, and traffic dynamics. The purpose of this research project is two-fold. First, we would like to contribute to one of the foundation aspects of the agent-based computational economics (ACE), known as agent engineering; second, from there, we would then extend our contribution to enhance the validity and robustness of market design. Market design is probably the most policy-oriented application of ACE. Some of its recent progresses, particularly in the electricity market, are so encouraging that they lend promising support for the future of ACE. However, it is also well noticed that two issues concerning market design have to be carefully addressed before further advancing: validity and robustness. These two issues are mainly due to different models of stakeholders. The way that we model the stakeholders` learning, planning and action may end up with different outcomes. Recent results on the agent-based electricity market, for example, already evidenced this point well. Nonetheless, this issue is generally shared among almost all agent-based models. Over the past decade, many Computational Intelligence (CI) methods have been applied to the design of autonomous agents, in particular, their adaptive schemes. This design issue is non-trivial since the chosen adaptive schemes usually have a profound impact on the generated system dynamics. The lesson, known as 「agent engineering matters」, is probably the most important one that we have leaned from the pile of the literature on agent-based computational economics in general, and on market design in particular. Robert Lucas has suggested using laboratories with human agents, also known as Experimental Economics, to help solve the selection issue. While this is a promising approach, laboratories used in the current experimental economics are not computationally equipped to meet the demands of the selection task. As well indicated in Herbert Simon`s study, different decision supports, associated with different computational power, may determine how agents learn. Therefore, this research project, as mainly motivated by Simon, attempts to materialize Lucas` suggestion by establishing a laboratory where human subjects are equipped with the computational power that satisfies the computational equivalence condition. The ultimate objective of this research project is to build up a laboratory in accordance with the proposed computational equivalence (CE) principle. The lab, called the CE Lab, can integrate agent-based computational economics and experimental economics into a coherent body so as to enrich our understanding of the markets comprising of both human agents and software agents, and provide a foundation of agent engineering. The CE Lab is built upon the extensions of two existing simulation platforms developed by this research team. One, from the side of agent-based computational economics, comprises the double-auction market software, AIE-DA, and the artificial stock market software, AIE-ASM. The other, from the side of experimental economics, is the on-line web-based experimental market, Taipex (Taiwan political future market). In the first year, we shall expand the current version of AIE-DA and AIE-ASM to allow for different trading mechanisms and behavioral modules. In the meantime, step by step, we will integrate software agents into Taipex and test its performance in the two forthcoming major domestic elections - the provincial city mayor`s election and the legislator`s election. When more and more software agents are introduced to Taipex, a tool-box from which human agents are able to choose their designated behavior will be established, and we shall begin to see the merge of the two platforms. This is expected to be done in the second year. In the last year, the CE Lab will be put into action by actually being tested on a number of selected important market design issues from the literature. During the three years of research, great experiences can be gained by maintaining a close contact with a Japanese research team led by Prof. Hiroshi Deguchi at Tokyo Institute of Technology, who is developing a very similar system, yet with limited capability in facilitating the study of agent engineering, known as U-Mart. |
