Hacker News new | ask | show | jobs
by seanica 4927 days ago
One thing that mustn't be over-looked with interfaces that 'learn about your behavior' is they can lock into a 'local maxima' and can be difficult to retrain without resetting to factory defaults. - If your lifestyle changes, can the interface keep up?
1 comments
jbrennan 4927 days ago
That could happen, but really that would be a "bug", not an inherent problem with the design. That would be the developer's job to fix.
link
seanica 4927 days ago
I disagree that this would be a bug. It's a design flaw that cannot be corrected by fixing bugs.

It's analogous to security flaws. If there is a flaw in the design, no amount of bug fixing will make the system secure, unless that 'bug fixing' changes the design.

link
jbrennan 4927 days ago
Can you explain why you think it's an inherent flaw in the design?

The way I look at it is, yes the software should keep a history of user behaviour and base its actions off that, but there must be feedback involved, either explicit or implicit. This way, if I gave some input to the system once but then never did so again, the likelyhood that one event should affect the future would diminish over time.

There could be trickiness around "Bubbles" (like a Search bubble, where it only recommends to you things it thinks you'd like, and never shows you other things). I think those are problematic and should be dealt with. But I don't think that means it's impossible to fix. It's just something that needs to be thought through. I don't have an answer for it right now but that doesn't mean there isn't an answer.

link
seanica 4927 days ago
> It's just something that needs to be thought through.

Your statement is what I mean. "Thinking things through" should be done during design. Once you have built system, its much harder to compensate for design flaws.

Programming is not designing. Designing is not programming. Fixing bugs is not designing.

You have to design into the UI system a means for it to compensate for changes in user behavior. You don't want a system that takes many uses to train. At the same time you don't want a system that is trained by a single use. For me this is the crux of the problem.

The happy medium that automatically detects deviations from a user's 'normal' behavior _and_ takes the correct action is very hard to design, as it involves AI fuzzy logic.

link
jbrennan 4927 days ago
>Your statement is what I mean. "Thinking things through" should be done during design. Once you have built system, its much harder to compensate for design flaws.

I agree with this fully. This is something that would need to be solved before building such a system.

A great example of a system which learns from history, but which also supports changes in behaviour is: http://worrydream.com/MagicInk/#engineering_inference_from_h...

The linked example is specific to one application, but he continues to detail what he thinks would make for a general solution.

link