Greenfoot Live

There are many aspects to making a learners’ programming tool successful. You obviously need the tool to be working and useful, but you also need to have material available on how to use it. Over the past year or two, our team been very busy with a lot of technical implementation work: having made our Stride editor available, we’re now neck-deep in a rewrite of BlueJ’s GUI from Swing to JavaFX. So, we have fallen a bit behind on our efforts to communicate how to teach with Greenfoot, and to engage with teachers who are using it. But this Monday we started a new initiative to try to rectify that: Greenfoot Live. Our plan is that every two weeks on a Monday at 17:00 UK time (which is UTC+1 over the summer), we’re going to do about 30-40 minutes live stream talking about Greenfoot. If you can make it live, great — but if you can’t then you can still watch the recording afterwards.

Our first show involved Michael and me covering how to display text on the screen in Greenfoot using the showText method, but we also discuss a few software design choices and encounter an exception along the way:

The next live stream will be on Monday 22nd May at 16:00 UTC, available via our Greenfoot Youtube channel.

What Do Novice Programmers Write Literally?

Recently I was asked about use of hex and floating point literals (especially “E” notation) in the Blackbox data set: do beginners use them? I was intrigued enough to knock up a simple program to find out. My method is quite straightforward: I take the latest version of each source file which successfully compiled, run it through a Java lexer and pick out the literals. This gives us about 40 million source files to look at.

Before we get into the results, here’s some predictions I made beforehand about our data (where most users are assumed to be programming novices):

  • Very few users use hex literals
  • Most hex literals are 0xFF or similar bitmasks
  • Almost no-one uses underscores (Java lets you write numbers with underscores, e.g. 1_000_000)
  • Almost no-one uses E notation (and when they do, mainly 1e-6 for epsilon values in floating point comparison)
  • Most floating point values are between 0 and 1

Hexadecimal Integers

Let’s start with hexadecimal integer literals. There were 814,920 hex integer literals, compared to 29,044,559 decimal integer literals. So 2.7% of hex/decimal integer literals were in hex. (I didn’t bother going into octal, but there were a handful of uses. I suspect many of these were an accident.). That is a bit higher than I was expecting, admittedly. In terms of their value, here’s the top five:

  • 0xFF: frequency 89,663
  • 0x0: frequency 52,732
  • 0x30: frequency 16,742
  • 0xF: frequency 16,009
  • 0x1: frequency 13,799

There are two F bitmask values there as predicted. I was a bit surprised by how many zeroes and ones were in there: why write them as hex (0x0) and not just decimal (0)? My guess is that they are working with bitmasks nearby, and out of habit/consistency write the values as hex.

Decimal Integers

There’s not too much to say about decimal integer literals, but I will mention the most frequent items. It’s a sequence that runs as you might expect (zero being most frequent, and increasing numbers being less frequent), punctuated by some numbers which testify to computing’s love of powers of two. Most frequent first:

0, 1, 2, 255, 128, 3, 4, 5, 256, 8, 10, 7, 6, 100, 127, 16, 20, 1000, 9, 50

The frequency is a decreasing power law (1 is half that of 0, 2 is half that of 1, then the tail begins to flatten out).

Underscores

Underscores are a relatively recent addition to Java (added in Java 7) and little-known. Indeed, only 692 decimal literals had underscores: 0.002% of all decimal literals. Oddly, 737 hex literals had underscores, which as a proportion is much higher: 0.09%. I suspect this is because both underscores and hex literals are both used by more advanced users. Generally though, our users are clearly not making much use of this underscore feature.

Decimal Floating Point

There were 1,791,915 floating point decimal literals. Of these, only 3,002 used the “E” notation (e.g. 1.15E12): 0.16%. Clearly not a very used feature. As for their values, the top five were: 1e-3, 1e-8, 1e-6, 1e6, 1e-20. I’d say my prediction about the use for epsilon values was borne out.

Regardless of notation, across all floating point decimal literals, the most frequent values were: 1.0, 0.0, 100.0, 3.0, 2.0. Technically, my prediction that most values were between zero and one was almost correct: 47% of values were between zero and one. But really, this is only because 23% of them were zero or one. As a last side note on these literals: 7,130 (0.40%) started with a dot (e.g. “.5”) — something we disbarred in Stride due to the awkwardness of parsing that in expressions. But actually we could have banned E notation (also a pain) with less immediate impact.

Hexadecimal Floating Point

If you even knew that hexadecimal floating point notation was a thing in Java, then give yourself a pat on the back. Added in Java 5, they look like “0x1.fe2p5”, where p takes the place of the usual “E” notation because E is of course a valid hex character. I only know about this because we have a parser in BlueJ, which does accept these. I found precisely four uses of this notation, which is probably more than expected.

Limitations

This is a pretty cursory look at literals with a fairly crude methodology. Note that although we only looked at the latest version of each source file, source files in Blackbox are not independent of each other (e.g. if a teacher gives out a project with a floating point literal, that will show up identically in each student’s copy). For example, the four hex float point literals were the same value, suggesting they are not independent. And on a related note, I’ve only looked at source files regardless of whether they come from the same user or not, so we’re only measuring source occurrences here, not the number of users who use a particular notation. But I think our N is high enough that individual users cannot tilt the statistics.

Naming abstractions

Naming things is important: names permit more precise and concise communication. We don’t say “the movable clickable pointer controller box”, we say “mouse”. In computing we face the challenge of naming a lot of things, both physical and virtual. Some names are intended to be useful analogies: files or documents are like their physical equivalent. This can quickly get quite tenuous: scrolling is named after the action needed to read a historical scroll despite most people having never scrolled a physical roll of paper, and a mouse was named due to its tail.

Computing has been surprisingly effective at digging up and re-using words which pre-date computing, and vastly increasing their use:

(What was delete used for before computing, I wonder? Ledgers?)

Naming difficulties

Sometimes we can find useful words to refer to computing concepts, even as the concepts become more abstract. A sequence of items is a list. An group without duplicates is a set. A list/set is a collection. That’s not too bad. It gets harder when you need a name for a data structure which corresponds one value to another. We usually talk about keys and values, and variously call this collection a “associative array”, “dictionary” or “map”. It’s clear that there is no useful existing word to borrow which carries the right meaning, so we must just pick one and collectively learn what it means.

A list is a type, as is a string or integer. But a list can have different inner types, so we want a different name for that: a polymorphic type. But a list (which has one inner type) is different from a map (which has two), and so we need a name to talk about different types of types. For this, type theory refers to kinds: the arity of types, where arity is the name for the number of parameters that something takes. If your brain is creaking at this point: is that because type, kind and arity are poorly chosen words, or just because the concepts themselves are abstract and difficult? Not to mention all the other abstractly-named programming terms: class, interface, protocol, metaclass, polymorphism — the list goes on.

I think people struggling to learn something with a non-obvious name (i.e. most programming concepts!) often make a false assumption: that the concept is difficult to learn because the name is unhelpful. (And thus if it was just named better, it would be easier to understand.)

Names can be helpful if they are descriptive (e.g. a network star topology) or transfer intuition (e.g. memory in computers). But intuition is not always available for abstract concepts, such as classes in object-oriented programming. As we get progressively more abstract, names aren’t going to save you from the difficulty of understanding an abstraction: the name isn’t the problem, the concept is.

Living Computing

SIGCSE this year was held in Seattle. It turns out there is a computer museum in Seattle, not too far south of the downtown area, called the Living Computer Museum. The key distinguishing feature is that almost all of the computers there are actually running, and usable by visitors to the museum. Sue Sentance, Ian Utting and I headed off to take a look.

The ground floor is filled with modern gadgets and sensors which were all too familiar after spending the week at a computing education conference: a real busman’s holiday. But on the first floor were all the vintage computers, which were much more interesting. There is an oscilloscope which is running Tennis for Two, considered to be one of the first computer games:

This typifies what is great about the museum: it uses original hardware, not emulation, and is usable by visitors: I was playing the game with Ian while recording the video.

Historic Computers

If there’s one lesson I’ve learned from computing museums, it’s that you either die by age 25, or live long enough to see your childhood computers in a museum. I was amused to find an Amstrad in the museum, given that it’s a British manufacturer. The Amstrad was a fairly cheap machine, and the company’s head Alan Sugar has always had a bit of a reputation as a producer of low-quality goods. Hence my amusement at this:

After a moment or two I worked out how to reset the machine:

And then I found Sim City in that box of 5.25″ floppy disks. I’m not sure I expected to ever again use a floppy disk! I first accidentally ejected the floppy with the OS on it, but unlike USB sticks which complain immediately if you yank them out, older PCs didn’t even notice if you didn’t do anything to access the disk while it’s out. So I managed to load Sim City:

To explain: older games used to prevent copying by having special codes in the manual which you had to enter on load. To avoid being defeated by photocopies, they were often printed in yellow-on-white or black-on-black, and/or spread throughout the manual so you’d have to copy the whole manual. I did google on my phone, but at the time I couldn’t find the relevant info (now found, for the curious).

Interface Evolution

There is a card-punching machine in the museum on which you can punch your own cards. I’m a bit too young to have done any card-punching myself, but it’s quite an experience. You can’t see the last 1-2 characters you typed so it’s easy to get lost punching even simple sequences. There’s no backspace/undo, of course. And the keyboard itself is a marvel of bizarre design:

Punch cards are quaint curiosities now, but what a bloody awful way to program. 10x programmers may not exist, but we are all 100x programmers compared to the punch card era.

I also found a computer running Windows 1.0, which I’d never used before (3.1 is where I came in). The basic ideas of GUIs are there, but it’s horribly slow and clunky:

Summary

If you’re ever in Seattle, I highly recommended visiting the Living Computer Museum. It’s not a huge place, but it is my favourite of the computing museums I’ve been to, for the simple fact of being able to use the computers. It’s something I’ve mentioned before when discussing computing research: I don’t think you can truly understand an interface without using it. You can’t appreciate how clunky Windows 1.0 was until you drag the horrible ball mouse across a mouse mat and have to hold and drag the mouse to access menus and mess around with monochrome Paint.

Finally, I’ll end with this, as a thought on how computing has evolved (the scale isn’t apparent here, but it’s about a metre tall and a couple of metres wide):

Stride and Git arrive in BlueJ preview

We’ve just released a preview version of the next major BlueJ release, 4.0.0. It’s version 4.0.0preview2, available for download from the main website. There’s several features available in this preview release:

Stride is our blocks-like structured code editor which we added to Greenfoot 3, and we’ve now added it to BlueJ as well. There’s lots of details elsewhere on Stride so I won’t reproduce it all here. There is our guide with a simple text+pictures overview of the editor. We recently noticed that a conference talk we gave at last year’s JavaOne made its way online, so you can watch a video of that on Youtube. And Michael has been making a few short videos about Stride over on his blog, which you can watch. We’ve now included two-way conversion from Stride to Java and Java to Stride, so it’s easy to take your existing Java projects and convert them to Stride to get a good feel for the editor.

stride-editor-circle

We’ve also added Git to BlueJ. People have been asking for Git support for years, but previously we only had Subversion and CVS(!) support. We’ve now added Git, with a fairly simple interface that makes it easy to get started with using Git to version control and share BlueJ projects. We’ve got a draft tutorial online for Git in BlueJ. (And CVS support has now been removed.)

Another major change is in the error highlighting. Previously, BlueJ would only show an error when you hit the compile button, and then only one. That’s now changed to match the behaviour of most Java IDEs: errors are shown as a red underline as you write the code, and if there are several errors, they are all highlighted. It will be interesting, once the full release is out, to look at our Blackbox data and see what effect this change has (if any) on programming behaviour in BlueJ.

bluej-errors

The last change is one of the most time-consuming but least catchy. We’ve rewritten large parts of BlueJ’s interface (from using Swing to using JavaFX, the newer Java GUI toolkit). Along the way we’ve improved various features, which I’ll talk more about another time. Probably the most noticeable change is that we now support tabbed editors (you can see the tabs in the pictures above), with multiple editor tabs in one window, rather than always having a new window for each editor (another oft-requested feature).

It’s called a preview release because we know it’s not quite finished: there’s still a bit of GUI to improve, some small bugs to iron out and so on. But we think it’s close enough to let everyone have a play with it. If you spot any issues, let us know: in the comments here, by email to support@bluej.org, or on the Blueroom teachers site.

Evaluating our Frame-Based Stride Editor

Last week at ICER, Thomas Price presented a paper: “Evaluation of a Frame-based Programming Editor”. I provided some technical assistance for the data recording aspect (and not very proficiently — I was responsible for some lost data), but Thomas and his local co-authors at North Carolina State Universiry (NCSU) did all the hard work.

The study looked to compare our frame-based Stride editor to text-based programming. Since Greenfoot supports both Java and Stride, it makes for a rather neat experiment design: you can get students doing exactly the same tasks in the same IDE, with only the editor component differing between the two conditions. Thomas & co at NCSU set up such an experiment, with one class of middle-school students doing a Greenfoot exercise in Java, and the other condition using Stride. We recorded detailed data on their programming activity by re-purposing the code from our Blackbox data recording project.

It’s no surprise to say that the Kent side of the team were hoping for positive results for Stride, as we are personally invested in our new tool. The nice thing about this study is that with the NCSU side doing all the data recording and all the analysis of the results, we managed to minimise any such bias. And as often happens in real experiments, the results do not have a single clear narrative. My one-paragraph gist is as follows:

The students in both conditions rated the activity as low frustration and high satisfaction. No differences in satisfaction were found between Java and Stride; there exists, however, potential of a ceiling effect. Students in the Stride condition advanced through the task instructions faster than the Java side and completed more objectives with less idle time than Java. Less time was spent making syntactic edits in Stride than in Java, and less time was spent in Stride with non-compilable code.

Mark Guzdial pointed out on twitter that it was quite surprising that since the Java students did seem to tail off, spending more time idle and completing less of the later objectives, that there was no different in frustration, satisfaction or performance:

I wonder if any difference in frustration/satisfaction may also be masked by the ceiling effect: if students in Stride were super-satisfied and unfrustrated at the extremes of the scale, a slight reduction on the Java side may not have been picked up. As for the lack of difference in time, I’m guessing that the time saved by Stride’s syntax benefits is being partially absorbed by struggles with the editor paradigm, such as learning the command keys or editing expressions.

The paper is freely available to read, so if you are interested, take a look.

Measuring Errors

One interesting technical aspect of the paper is to do with measurement of errors. One of our hypotheses was that Stride should help to eliminate or minimise a lot of errors in Java, such as syntax errors (e.g. mismatched brackets) and forgetting method parameters.

In Greenfoot 2.x and BlueJ 3.x, students must explicitly click the compile button to compile their code, which either succeeds or displays an error. With explicit compilation you get information from the fact that the user clicked compile: they probably think their code is in a state where a compilation should succeed. And if in two conditions users compile with a similar frequency, you can compare the proportions of successful compilations to get an idea of any differences in error rates.

Greenfoot 3.x and BlueJ 4.x have changed this. Now, errors are delivered automatically, whenever you stop typing for a few seconds. So we lose the information on whether students think the code is error-free, and we get a lot of spurious errors. Worse, the error rates will differ arbitrarily between Java and Stride. For example, if you slowly enter “if (x < 0) {x=1+2;}" in Java, you'll get syntax errors from the moment you start typing until you enter the final closing curly bracket. In Stride, the code will be valid for a few intermediate states (e.g. the if-frame will compile successfully once you've entered the condition, but before you've begun the body). It is not fair in that case to say that Stride causes less errors.

Thomas and I struggled to find a way to sensibly compare error rates. Obviously there are other measures which we could and did use, such as task progress, or occurrence of a specific kind of error (e.g. mismatched brackets). But it is still interesting to compare programming error rates. The measure we settled on is "time spent with uncompilable code". It is not perfect, but it seemed like a reasonable first stab and better than say number of errors per minute, or percentage of successful compilations. I'd be interested to know of any other such measures, perhaps from studies of professional programmers where IDEs also automatically provide errors. The result seemed to favour Stride:

The Stride group spent on average 7.45 minutes less time with non-compilable code than the Java group, which spent on average almost half of the activity with non-compilable code.

I was also interested in another of the paper’s minor results. Our Stride editor provides customised help in a number of cases. If you begin to write a variable declaration in Java and you get as far as “int” then go off to look up something else (forgetting the variable name), you won’t get a very helpful error message from the Java compiler (“Not a statement”). In Stride, we can give the more specific error “Variable name cannot be blank”. We have a small number of these customised errors, but we found that four out of the top five errors in Stride were our customised versions. This means that Stride can replace the often-unhelpful Java compiler error messages in the most frequent cases, which I think bodes well for assisting novices.

Our current plan is to add Stride to the forthcoming BlueJ 4.x, and to add Blackbox support for recording it, which should give us some interesting data on the use of the language. (It will also provide an interesting opportunity to contrast behaviour between BlueJ 3.x, with explicit compilation, and behaviour in BlueJ 4.x, with automatic error display.) If you have signed up for access to Blackbox, you will also get access to all this data.

A case for publishing research software

A major part of research is acquiring and sharing knowledge. This manages not to be as straightforward as it should for political/business reasons (see: journal publishers, paywalls and open access), but technically it is at least simple. You write a paper consisting of words and pictures, other people download them and read them. Knowledge has been transmitted. Where life gets much more difficult is a newer but fast-growing part of research: sharing software.

One use of software in research is as a tool for doing analysis. This affects all the natural sciences, and there are issues with how to gain credit for producing software (see the new proposed policy on software citation, and the new journal of open source software). But within computer science there is an additional research role for software: sometimes the software is part of the research output. Nowhere is this more apparent than in areas like human-computer interaction research.

Publishing software interface research

The typical form of a modern paper about a new software interface is to provide a description of the interface, followed by an evaluation of the interface with human participants. Thus the research output is two-fold: design and science. Putting these both into a paper might seems sufficient.

The software research process.
The classic software research process. The researchers follow the process on the left, but everyone else only sees the output on the right. The researchers must have the software, but they are not required to share it, only to describe it in the paper.

However, accurately describing an interface design in text is a difficult task — the medium is just ill-suited. (Much like writing about music being compared to dancing about architecture.) It is difficult to describe the function of all interactions with the system: you’d write an endless series of “when the user presses left here they return to the home screen”; something almost akin to the original program code. You’d also need to describe not only the intended interactions but what happens when the user does something wrong. You also can’t use pure text: images are surely necessary to portray an interface. And that’s not to mention emergent properties which affect software’s usability, like the speed of the interface. Ultimately, if you want to understand the design of a software interface, there’s very little substitute for just using the interface.

Research Software Archaeology

Recently, I needed to write a detailed related work section for our work on frame-based editing. One of the challenges of publishing this work is that it is similar to work on the structured editors of the 1980s, which have largely failed to catch on. And additionally, it seems every reviewer knows a different editor, so each one seems to come back with “how is this different to structured editor X that I used in the 1980s?” [1]

So I end up searching for details about the design of 1980s structured editors. If there’s no paper and no software, there’s not really any way to find out about the design. If there is a paper, I hope that it has a reasonably detailed description of the editor (for example, the write-up of the Cornell Program Synthesizer). Regardless, I also try to search for a runnable version of the software. Ha!

There are few editors from that period which are available to run on a modern machine. Some were simply never released, partly because pre-Internet, sharing software was awkward. Some are unavailable due to their age: many of the structured editors were designed for processors or operating systems which are no longer available. So some editors seem to be totally lost — I can’t find any leads on downloading a copy of the Cornell Program Synthesizer, for example. Some other editors have a tantalising binary distribution which often cannot be run: for example, Boxer’s Mac binary.

The ACSE/GENIE editor, alive and running.
The ACSE/GENIE editor, alive and running.

I did have one or two successes, such as getting a version of the GENIE editor running in an emulator. And it was a revelation that greatly pushed forward my understanding of old structured editors. By modern standards, they were awful. The papers’ descriptions didn’t make clear how tedious and fiddly the navigation was, how unhelpful the editor was, how awkward it was to deal with errors. Running the software was an absolutely crucial step to comparing our work to theirs. It allowed me to understand the design and critique the editor’s operation for myself, rather than relying on the authors’ incomplete descriptions of their own software.

For all the other editors which I couldn’t run, there are these reviewers asking the perfectly valid question in research: “How does your work relate to previous work X?” And the honest answer is: I don’t know. Perhaps nobody can know any more — the paper wasn’t very detailed and the software is lost in time. This is no way to do research.

The Solution

The solution to all of this is readily apparent: if your software is part of the research output, you must publish the software. And a binary is insufficient; binaries too easily bit rot, refusing to run on modern systems with no way to fix them. Source code is what is needed.

This week Andy Ko made available his Citrus/Barista structured editor from the 2000s. I downloaded and ran it: the binary did run, but it spat out repeated exceptions and I wasn’t sure if that was impairing the functioning of the software. Thankfully, Ko didn’t just publish a binary: he published the source code. For this, I salute him. I went to modify the source code and it turned out not to compile with a modern Java compiler. After some tweaks I got it compiling, and then fixed the exception. Because the code was on github, one accepted pull request later and the software in his repository will now compile and run on a modern machine. This — this is how software research should be.

Published source code for software is crucial to allow later researchers to use, evaluate and compare the software. I fully understand that everyone feels antsy about publishing source code. If I’m honest, the Citrus source code is a bit confusing, somewhat lacking in documentation and the software seems a little rickety. But that’s how research software usually is; my research code for our Blackbox work is the same. I’m not particularly proud of that code, but my recognition that sharing the source is important marginally outstripped my embarassment at its quality. Research code will almost always be shaky and iffy [2]. It’s usually written by a single person (often not a professional software developer) for a single purpose, so it’s likely to be hacky and not well documented. Let’s all accept that research code is bad, and agree to share anyway.


Footnotes

[1] It’s interesting to note that when the researchers ask how our work compares, they are implicitly asking about the design, not the science. Given that almost all venues will only accept science or design+science, it’s curious that most of the comparison to related work is about comparing the design. This is at least partly because the science quickly outdates in software interfaces. Even if the older editor papers had performed rigorous evaluations (which they almost exclusively did not), the results don’t necessarily persist. If someone told you that editor X had been evaluated as easy to use and as good as text editors, tested on a 25-line text terminal on a 1980s thin-client Unix machine, would you say that was useful in evaluating editor X against a modern editor? Would it even be worth comparing the usability of our editor directly against a 1980s editor? I doubt it; the usefulness of the previous work is more in comparing our design to theirs, not so much our scientific evaluation against theirs.

[2] Given that we make software — BlueJ and Greenfoot — which we encourage people to use, I should point out that they are actually stable, reliable, and fairly well engineered! And open source, to boot. The setup of our research group and funding allows us to do this, making us blessed compared to other researchers. Quality software in research is of course possible, and the preferred option, but we must recognise that it is a rarity, and not let that get in the way of sharing.