HTTP/1.1 200 OK Date: Sat, 11 Oct 2025 01:24:22 GMT Server: Apache Last-Modified: Mon, 14 Apr 2003 11:42:54 GMT Accept-Ranges: bytes Vary: Accept-Encoding Content-Encoding: gzip Content-Length: 1789 Content-Type: text/html [This page is https://www.cse.lehigh.edu/~brian/pubs/1997/hci/]

Toward An Adaptive Command Line Interface

Uncorrected Paper (as published, 4 pages)
Postscript (247KB) PDF (124KB)
Brian D. Davison and Haym Hirsh

Corrected Abstract (see below)

This paper explores different mechanisms for predicting the next command to be used for the UNIX command-line shell. We have collected command histories from 77 people, and have calculated the predictive accuracy for each of five methods over this dataset. The algorithm with the best performance has an average online predictive accuracy of up to 38%.

Presented at the Seventh International Conference on Human-Computer Interaction, August 24-29, 1997, San Francisco, CA. Proceedings published as Advances in Human Factors/Ergonomics Volume 21B - Design of Computing Systems: Social and Ergonomic Considerations, pp. 505-508: Elsevier Science Publishers.

Important Note:

This paper, and our Autonomous Agents paper that it references have results that were computed using a corrupted datafile. While it does not change the results qualitatively, it does make some quantitative changes. Corrections to the paper are as follows:

• 3. Experimental Results

  In the real world, user productivity is the primary concern. Since the data collected was from unmodified, non-predictive command shells, we can only speculate as to the improvements in productivity afforded by the methods presented here. However, in the best case C4.5's predictions were correct up to 38% of the time. Recall that the lengths of the histories on average were 2184 commands, and the average command was 3.77 characters long. Assuming that a correct prediction could be inserted with a single character, and that commands recorded in the history were all typed explicitly (not using any historical shortcuts or command completion), this method would have saved just under 27% of the keystrokes typed, which is very close to the 28% expected if predictability were independent of command length.

  Figure 1: Macroaverage performance vs. training-window size	Table 1: Online predictive accuracy (for training-window size 1000)
  	algorithm macroaverage microaverage Omniscient 91% 96% C4.5 38% 37% Prefix 33% 30% MFC 21% 18% MRC 19% 22%

• 5. Summary
  This paper has described a number of methods for predicting a user's next command in a UNIX command-line shell. We found that relatively straightforward, knowledgefree methods were able to correctly predict the next command that the user would execute up to 38% of the time, potentially reducing the keystrokes needed by close to 27%. The best performance was achieved using C4.5 to learn prediction rules based on two previous commands, and a training window of either 500 or 1000 past commands.

Back to Brian Davison's publications

---

Last modified: November 4, 1997 Brian D. Davison