The IDA'01 Robot Data Challenge

The IDA01 conference featured a Data Analysis Challenge, to which all conference participants could respond. The challenge was organized around categorical time series data. A series of vectors of binary data was generated by the perceptual system of a mobile robot, and series of characters was taken from George Orwell's book 1984. In both cases, the boundaries between meaningful units (activities in the robot data, words in the Orwell data) were absent, and part of the challenge involved inducing these boundaries. More speci cally, in each case we suspect a time series contains several patterns (where a pattern is a structure in the data that is observed, completely or partially, more than once) but we do not know the pattern boundaries, the number of patterns, or the structure of patterns. We suspect that at least some patterns are similar, but perhaps no two are identical. Finally, we suspect that patterns have a hierarchical structure in the sense that shorter patterns can be nested inside longer ones. The challenge is to nd the patterns and elucidate their structure. A supervised approach to the problem might involve learning to recognize patterns given known examples of patterns, however, this challenge encourages unsupervised solutions, those in which algorithms have no information speci c to the data, other than the data itself. One reason for this stringent requirement is to see whether domain-general solutions will be developed. This is also the reason for providing two, quite di erent datasets: One hopes that an unsupervised patternnding algorithm that works on both data sets will provide some insights about general characteristics of patterns.