5.2. Test data

Creating quality test data for a question is a problem that is often underestimated. The test data should be designed in such a way that it tests both the efficiency and the correctness of the algorithm used, in such a way that semi-efficient or mostly correct algorithms will get a reasonable part score.

It is common to order the test cases from the most easy to the most difficult, but it is not always straightforward to decide what is easy so this isn't a fixed rule. However, generally the first test case should be no harder than the sample case, and should be solvable even by a moronicly slow but correct algorithm. Later cases should get progressively larger or more difficult until the last few test cases can only be solved by one of a few algorithms that are reasonably well implemented (you shouldn't require people to spend hours tweaking their algorithms to make them perfectly efficient, however). If there is only one variable controlling size, then it will usually grow exponentially over the test cases.

The easiest way to generate large data sets is to randomly generate test data. This is not always a good idea, since many algorithms perform much better on random data than on extreme cases, and so will get more points than they deserve (perhaps even a full score). You should create a mix of random and hand-designed test cases. Hand-designed could mean writing a program to produce a specific test case. In some cases it is worth adding slight perturbations to the pathological cases, so that they cannot immediately be detected as such (e.g. if there is one test case that is obviously the worst possible case, do not use it as a clever contestant may modify his program to detect it and hard-code the correct answer).

Apart from hand-crafted pathological test cases, try to differentiate random cases in some way. For example, in a graph theory problem, put in dense graphs and sparse graphs, connected graphs and graphs with many components, and so on.

At some point before a problem is used in a contest, it is vital to validate the test data, as it is surprisingly easy to accidentally violate the constraints of the problem statement or even fail to follow the input format. The USACO even goes as far as to have an update field for a validator (written in Perl) to be uploaded, and a problem is not considered ready for use until the test data is validated. The SACO does not yet have such a system, but a simple way to pick up most errors is to use assert within your model solution to check all the constraints. If your model solution gets a full score, you can be reasonably certain that your test data is valid, although it will not pick up more subtle errors such as trailing garbage or inappropriate whitespace.


[Prev] [Next] [Up]

Last updated Mon Jun 19 18:09:59.0000000000 2006. Copyright Bruce Merry (bmerry '@' gmail dot. com).