Artificial Intelligence Depot
Visiting guest. Why not sign in?
News, knowledge and discussion for the AI enthusiast.
FEATURES COMMUNITY KNOWLEDGE SEARCH  
Developers, get involved!
The AI Foundry promotes the development of Open Source artificial intelligence projects. Why not join it now? All projects welcome extra help!
Visit the AI Foundry
Experiment in Curiosity
 
Experiment in Curiosity

As part of my on-going research into computational creativity and the nature of curiosity I've created an on-line experiment where a person and a "curious design assistant" simultaneously use an interactive evolutionary tool to evolve two-dimensional (Latham-esque) forms called "horns".

The experiment has been broken into two parts, in the first part the "curious design assistant" tries to predict the preferred selections of the person, in the second part the tables are turned and the person is asked to predict the selections of the assistant.

Each part of the experiment should take no more than 10 minutes to complete. So if you have a little spare time or would just like to know more about the experiment drop by my webpage:

http://www.arch.usyd.edu.au/~rob/java/applets/horn_expt/index.html

Let me know what you think.

-RoB-

47 posts.
Tuesday 06 August, 07:17
Reply
Goals

I started playing around and after 25 clicks, I still wasn't sure what I was doing... Am I "guiding" the system to attempt to create some image I want (like you mention in the intro), or simply picking my favourite images for the sake of it?

I guess you mean the second, but that's not immediately clear by the description. I also think that 100 is a bit too large a number to reach!

That aside, I think it's amazing the amount of cool patterns that this thing can create! You'd expect to get bored of them after a while, but no...

What you going to use all the data for?

934 posts.
Tuesday 06 August, 15:29
Reply
Re: Goals (long reply)

Alex,

Thank you for taking the time to have a look at the experiment.

Your question is a good one. In general, interactive evolutionary systems are designed to provide a convenient tool for exploring a space of possibilities, they are less useful when you are trying to achieve something specific.

Therefore I would say that your goal is to try to find forms that are 'cool'. The instructions are deliberately a little vague on this point because I don't want people to feel like there is a "right way" or a "wrong way" to use the system.

Perhaps I should amend the instructions to emphasise that there are no right answers when trying to explore the space of forms. I have had others comment that they think they might be spending too long on each decision. 100 generations doesn't take very long to get through when you don't agonise over every decision -- it should take only 10 minutes or so to make 100 selections.

I think generative art systems like this one are great because they can offer so much variety with such little effort on the part of the user. It's slightly more difficult to get the tool to work correctly -- I've been working on this system so much recently that all I can see when I close are these images. :-O

The goal of the experiment is to compare the model of curiosity that I have developed with the exploratory behaviour of people using the generative art system. My previous pilot study indicated that the agent can do quite well and I'm hoping to confirm this result with the data I'm collecting in this study.

One of the principle criticisms of interactive evolutionary systems, such as Kai's Texture Explorer, has been that the user doesn't feel in control; with a low mutation rate the resulting forms often look alike and are of little interest, raising the mutation rate results in forms that seem to bear no relation to their selected parent(s). You may find that this is the case in certain areas of the design space presented with the experimental applet.

My goals with my research are to try produce better generative systems that can be used to filter out uninteresting examples of a style, and provide an "autopilot" that can explore the space for you and bring back the most interesting of its finds.

Used as a kind of filter, a curious design assistant could ensure that the forms presented are interesting without the user having to play with the mutation rate. Obviously, in addition to modelling curious behaviour the tool would benefit from the ability to build a profile of a user's other aesthetic preferences but my current research is focussed on the modelling of curiosity in the creative process because there has been hardly any computational work in this area.

A curious design agent acting as an "autopilot" provides opportunities for creating non-interactive versions of generative systems that can produce a continuous stream of interesting works within a given style without user intervention. This possibility has some potentially interesting commercial applications which I am investigating now.

This may sound a lot like user profiling or data mining and indeed it is. The idea that novelty is "interesting" is one that has been investigated in data-mining systems and this work is built on similar ideas applied to creative domains like art and design. Potentially similar sorts of curious agents could be used to find "interesting" images or documents on the web. I might have a go at that in the near future but for the moment I'm concentrating on generative processes.

Ooo... that was a long reply. Well, you did ask! ;-) I hope other people will be as interested in taking part in this experiment.

-RoB-

47 posts.
Tuesday 06 August, 19:26
Reply
More Pretty Patterns

As a reward for those people who managed to get through reading my long reply to a short question, here is a another applet that I've just finished writing:

http://www.arch.usyd.edu.au/~rob/java/applets/fractals/JuliaNebula.html

This applet provides an endless stream of randomly generated strange attractors. There's nothing smart about this applet, there's no curious agent behind the curtain pulling the fractal strings, but it does produce some beautiful patterns.

If you're tired from performing those 200+ mouse clicks whilst doing my experiment then follow the above link, expand the window to full screen (press F11 on IE), stick some chill music on the stereo, turn the lights off and relax... ;-)

Thanks to everyone taking part in the experiment,

-RoB-

47 posts.
Tuesday 06 August, 22:09
Reply
Re: Announcement

Alex,

Thanks for putting this little announcement up on the front page. I am very surprised to see it make the news. ;-)

Erm... I just want to point out that it is not the curious design agent that generates the forms, in fact it is a general-purpose agent that uses a generative art system in the same way that a person using the same system would. I've actually used the same agent architecture (but with a slightly simpler vision system) to evolve Karl Sims' style images -- see my papers on "artificial creativity" on my academic home page for more information.

Personally I find the idea of an agent using a generative art system in the same way as a person more exciting than if I had written an agent only to produce these images. Maybe one day I'll be able to incorporate some of this stuff into Fear and make bots that stop mid-fragging to paint a picture? ;-)

Sorry to be a pedant but you know how it is when it comes to your academic research... :-/

-RoB-

47 posts.
Tuesday 06 August, 19:59
Reply
Curiosity killed the cat

I quite enjoyed doing your experiment, managed to take a look at some interesting designs! It would be interesting to be able to look at the results afterwards though, I felt a little cheated when I finished and realised that the results weren't displayed at the end.

What technique are you using for the user modelling?

I'm assuming curiosity agents can be applied to anything that involves a measure of human subjectivity. If an agent can learn all manner of likes and dislikes of the subject then it can work as an information gathering agent - music, intellectual interests, news, art, the list is endless. The web is the ideal ground for something like this to opperate. It could also be a handy marketing tool - agents that actually learn what a user likes and wants: the ultimate market research tool!

Linden

26 posts.
Wednesday 07 August, 09:20
Reply
User Modelling? Not me...

Linden,

Thank you for taking the time to do the experiment. I'm glad you enjoyed exploring the space. In the future I'll be creating more general interactive evolutionary tools to explore other design spaces. I'll be sure to post any announcements here.

I'm sorry you felt cheated at the end of the experiment but you really don't have to. The truth is that you could see the results for both parts of the experiment in the bar graph. This shows how much the agent agrees or disagrees with your choices. A peak in the graph on the left-hand side means that you and the agree generally agree. I'll write some more about the significance of the results once the data is collected.

I can understand your confusion about not seeing the "results". Perhaps you were expecting to see a collection of images that would somehow capture your preferences? As one might expect with user modelling. However, I think the really cool thing about this system is that I'm not doing user modelling: I'm actually comparing a computational model of curiosity based on theory and previous empirical study with the behaviour of people using a generative art tool to see how well they match. In other words, in this experiment I'm comparing the behaviour of an Artificial Intelligence with a Real Intelligence -- in a very limited domain!

The surprising thing is that the model tends to a pretty good job of predicting people's selections even without user modelling in the more traditional sense.

You're not the first person to assume that what I am trying to do is capture a set of user preferences. In the future I hope to combine user modelling with the model of curiosity to provide a more accurate predictor of a user's selections -- this would be useful if people were trying to use a curious assistant to design something with a particular aesthetic not prescribed by the generative system.

The problem with using traditional forms of user modelling with an art or design system like this one is that they tend to learn the "average" of a person's likes and dislikes but when a person is exploring a space they don't want to see the same things again and again. In some sense a curious agent does exactly the opposite of this type of user modelling because it is always looking for designs that are different and therefore more "interesting". Instead of trying to model a person's preferences for design products I am using a model of a person's design process within the limits of the tool.

I should probably add a section to the background information section of the introduction to the experiment to say that I'm not doing user modelling.

I'm glad that you can see some potentially exciting applications for curious agents. I'm certainly very excited about what these agents could do when partnered with conventional user modelling techniques. When I combine a curious agent with a user modelling system I will want the user modelling system to learn those qualities of selected works that are constant -- i.e. to learn prototypes for selections that cannot be explained by the model of curiosity.

I plan to write an article about my study of curiosity and creativity for AI-Depot after I got my current batch of papers finished. :-)

-RoB-

47 posts.
Wednesday 07 August, 10:46
Reply