機器創造力：創作的知識輸入～～在人工智慧、認知科學與哲學領域中，關於機器創造力的討論已經相當多。然而，直到最近，足夠數量的創意程式才真正出現，使我們能從中歸納出一些共通性。這些共通性讓我們得以提出一些方法，用以設計、運用並評估創意程式。在這個過程中，一個重要的部分便是確立一組可用來估計程式創造力的衡量指標。

在某些情況下，這些指標也許可以用來判斷某一個程式是否比另一個更具創造力。然而，此類情況必須十分特殊，並需考慮兩個程式的設計、輸入與輸出，而即便如此，這樣的比較仍很可能被認為不公平。類似地，也許可以利用創造力指標來判定某程式是否具有任何創造力，但這同樣存在問題，因為「創造力」是一個負載極重且高度主觀的詞。例如，Cohen 的 AARON 程式（Cohen 1995）被許多人視為具有創造力，但其作者本人卻不這麼認為。

因此，創造力衡量指標的價值主要在於：用於創意程式的設計，而非用於評估已既有的程式。如果我們能就一組程式創造力的衡量指標達成共識，那麼撰寫程式的人便可將其作為提高系統創造力的指引。倘若某程式的新版本在這些指標下的評價比前一版本更佳，那麼便很可能代表已有進展。

對於任何聲稱具有創造力的程式，一個重要的問題是：它的設計（包含其使用的演算法與資料）在多大程度上是被刻意安排以產生特定輸出。也就是說，程式是否被「微調」來生成特定結果？微調的證據會影響我們對它創造力的看法。例如，Lenat 的 AM 程式（Lenat 1982）之所以顯得具有創造力，其中一個原因是它最初在集合論領域工作，但後來轉向數論，而其最佳結果也來自後者。然而，當（Ritchie & Hanna 1984）顯示 AM 將「袋子（bags）」視為「數字」——這一導致它探索數論的「決定」——其實源於一些精心設計的知識時，我們願意接受 AM 為創造性的程度便因此降低。

為了探討輸入給程式的知識如何影響其創造力，我們首先討論適用於我們創造力衡量方式的程式種類，並提供一些此類程式的例子。接著，我們借用波普（Popper）的科學哲學來作為研究微調問題的理論基礎。之後，我們概述 Ritchie 先前在衡量創造力方面的工作（Ritchie 2001），並說明我們將如何在此基礎上延伸。接著，我們引入「創意集合（creative set）」的概念，並以此推導出一些與微調相關、且會影響程式創造力評估的衡量方法。最後，我們以 AM 與 HR 兩個數學理論生成程式（Lenat 1982；Colton 2001）作為案例研究。

There has been much debate about machine creativity in artificial intelligence, cognitive science and philosophy. However, only recently have sufficient numbers of creative programs become available for us derive some commonalities between them. Such commonalities allow us to suggest ways in which creative programs can be designed, utilised and assessed. An important part of this process is the determination of a set of measures which can be used to estimate the creativity of a program. Under certain circumstances, it may be possible to use such measures to determine whether one program is more creative than another.

However, the circumstances would have to be very special, taking the design, input and output of both programs into consideration, and it is still likely that such a comparison would be deemed unfair.

Similarly, it may be possible to use measures of creativity to determine whether a program is being creative at all, but this is also problematic, because creativity is such an overloaded and highly subjective word. For instance, Cohen's AARON program(Cohen1995)1 is cited as creative by many people, but is not thought of as creative by its author.

The worth of measures of creativity therefore lies in using them in the design of creative programs rather than the assessment of established programs. If we can agree upon a set of measures of creativity of programs, then someone writing a program can use these as a guideline for increas ing the creativity of the program. If a new version of the program is assessed more favourably by some of the measures than the previous version, it is likely that progress has been made.

For any program which purports to be creative, an important question is the extent to which its design (including the algorithms and data it uses) is contrived to produce partic ular outputs.

That is, has the program been `fine-tuned' to generate specific results? Evidence of fine-tuning can affect our perception of how creative a program has been. For in stance, one of the reasons that Lenat's AM program (Lenat 1982) appears creative is that it began by working in set the ory, but switched to number theory, with its best results arising in the latter domain.

However, our willingness to accept AM as creative is lowered by the evidence provided in (Ritchie & Hanna 1984) that AM's `decision' to regard bags as numbers — which led it to investigate number theory — was the result of certain carefully crafted knowledge.

In order to address the question of how the knowledge in put to a program affects its creativity, we first discuss the nature of the kinds of program to which our measures of creativity apply and give some examples of such programs.

Wethen draw on Popper's philosophy of science to motivate our investigation into fine-tuning. Following this, we give an overview of Ritchie's previous work on measuring creativity (Ritchie 2001), and discuss how we will build on these.

Wethen introduce the notion of a creative set and use this to derive some measures of fine-tuning which affect the assessment of creativity in programs. Finally, we perform a case study using the AM and HR mathematical theory formation programs (Lenat 1982), (Colton 2001).

The Effect of Input Knowledge on Creativity by Simon Colton, Alison Pease, Graeme Ritchie fr. Informatics Research Report EDI-INF-RR-0055. DIVISION of INFORMATICS, Centre for Intelligent Systems and their Applications, Institute for Communicating and Collaborative Systems November 2001， Appears in Proceedings of ICCBR-2001.