web interface test

[Starglider]

Firefox users, does this work for you?

interface prototype

Good for gridfont design and viewing (see 'load' button) only at the moment, the AI part (which generates a whole font based on a few characters to sample the style) isn't working yet. Does not work in IE because IE suffers from several concurrent forms of brain damage, which I have not yet been able to compensate for.

[Bounty]

The interface works for me (Firefox 3.5, Win7). One thing I'd change is the behaviour if you click an intersection of two diagonal lines: could you make it so that clicking it marks/unmarks a cross, rather than just one of the diagonals?

[Starglider]

The interface works for me (Firefox 3.5, Win7).

Thanks.

could you make it so that clicking it marks/unmarks a cross, rather than just one of the diagonals?

That is a bit fiddly to implement, but I will put it on the list of things to add later.

[Thanas]

It works for me to - Firefox 3.5.2, Windows Vista Enterprise.

[Vendetta]

Works for me, Firefox 3.5.5, OSX 10.5

[Starglider]

Incidentally I could use some volunteers to make fonts for this project, since the engine works better if it has lots of material for training. Also, scoring the output of the AI, once it is generating characters correctly. If you're interested, PM me for details.

[TimothyC]

Fx 3.5.5 on XP sp3, the Regenerate button does not work.

[Edi]

Works fine on FF 3.0.9, which is what I have at work. Can test on FF3.5.5 at home.

Looks awesome, by the way.

Will it stay available online later?

[Starglider]

Will it stay available online later?

It should be up most of the time. I only usually take web demos offline for a minute or two when I sync the server with my development workstation.

The next version will color the characters red/yellow/green, based on whether the engine could not recognise them, could recognise them only in the context of the rest of the font, or could recognise them context-free.

I am planning a public release of a 'Landmarks in AI' series, consisting of three browser-based upgraded remakes of classic AI programs, all running on our engine. This is #2, the first one is a remake of SHRDLU (which is nearly finished), the third one will probably be a remake of Sonja/Amazon (by David Chapman, as detailed in 'Vision, Instruction and Action'). Timescale depends on how busy I am with commercial work, but at the current rate, probably another six months or so.

Reasons, firstly the history of AI research is under-appreciated (loads of newcomers ignore it entirely and thus repeat basic mistakes), secondly the field as a whole has very little of the independent replication which a real scientific or engineering field should have (in particular, I find it ridiculous that people claim to have the secret to general AI, but their engine can't even replicate the state of the art as it existed in 1970).

[Edi]

Thank you. This thread has now officially crossed the threshold from Testing material to permanently worthwhile status and is thus punted to G&C.

[Rogue 9]

Works for me. Firefox 3.5.5, WinXP Media Center edition.

[FSTargetDrone]

Works for me with FF 3.5.5 in Win XP Pro SP3 (and in Opera 9.64 as well). The page in Opera seems to load a bit more slowly--each character on the right side loads one-by-one, in contrast to Firefox where they all seem to pop up at the same time.

Conversely, there is a slight delay in the drawing of the character segments as a new character is chosen in Firefox, but the segments with a chosen character in Opera appear instantly.

Not knowing much about any of this, Opera seems to be caching everything (which might account for the initial delay as the characters are drawn in the table on the right) at the outset. Then the segments for each character pop up instantly after the page is done loading. This is not the case with FF, evidently.

[Starglider]

Works for me with FF 3.5.5 in Win XP Pro SP3 (and in Opera 9.64 as well). The page in Opera seems to load a bit more slowly--each character on the right side loads one-by-one, in contrast to Firefox where they all seem to pop up at the same time.

Conversely, there is a slight delay in the drawing of the character segments as a new character is chosen in Firefox, but the segments with a chosen character in Opera appear instantly.

The difference is most likely due to the way the image composition works in the two browsers. This page uses a large number of translucent PNGs which are hidden, shown and retargetted as required. Most likely Opera has more efficient rendering code (Firefox suffers a bit due to excessive platformisation and portability issues) but less efficient layout code and DOM tree manipulation functions. The whole thing could use a code review by a full-time web app developer; I have basic competence at making web apps but client-side javascript really isn't my focus.

[The_Saint]

Loads fast and fine in Opera 10 (in Vista) and no apparent issues.

[D.Turtle]

What is the Regenerate buttonsupposed to do?

It doesn't do anything that I can see for me (Firefox 3.5.5, Windows XP).

[Starglider]

What is the Regenerate button supposed to do?

It is supposed to do two things;
1) Highlight all the human-produced characters depending on whether the AI could recognise them as their target letter or not.
2) Regenerate all the AI-produced characters to reflect the AI's best guess at how to extrapolate the style of the human-produced characters.

Neither of these is working yet. Right now I am tweaking the virtual coding scheme for geometric abstractions, so that the base prior for letterform construction is better aligned with how human perception of letters actually works. Strictly speaking, this is not entirely necessary, because after being shown enough training fonts (20 or so) Bayesian updating of the probability distributions over geometric breakdowns for individual letters will backpropagate and generate a reasonable domain-specific overlay for the base prior. In theory anyway, I will be doing a test sequence to check this later. However at the system's current stage of development it is better to hand-optimise the geometry concept group, because then I can be confident about it generalising correctly (paradoxical, eh?). This is similar in some ways to the original Letter Spirit, which had hand-optimised codelet chains and link weights in its semantic network. However the theoretical framework is quite different; I am using formal information and probability theory, the original was completely informal and designed in an 'intuitive' manner.

I was going to start a proper thread once the demo is running properly, but since this one got promoted from Testing, I guess I will post here when there is something new for people to try.

[Sarevok]

This is astounding. Once refined and improved such a software could automate a lot of grunt work. A version specialised for makinf simple buttons or icons alone would be a huge life saver.

Do you have any for taking this any further ?