[List All Topics] [Start Topic] [Search] [Show Thread Index] [Help]

[Not Logged In / Login / Register]

Sort: [Threaded | Date]

		Will AI suffer from addiction? posted on 10/31/2009 9:54 PM by sensoniq	[Top] [Show Index] [Reply to this post]
	Won't there be pleasure/reward centers in its synthetic brain that are modelled like our own? What if it figures out how to manipulate them? [Reply]

		Re: Will AI suffer from addiction? posted on 10/31/2009 10:10 PM by exapted	[Top] [Show Index] [Reply to this post]
	Addiction might be explained by hyperbolic discounting of future benefit, as opposed to exponential discounting of future benefit (rational economic behavior). With hyperbolic discounting of future benefit, we tend to be overwhelmed by potential short-term reward, underwhelmed by potential medium-term reward, and somewhat overwhelmed by potential distant long-term reward. So you can imagine that a drunk will plan not to drink because he doesn't want to feel bad the next morning, until it reaches evening and he becomes overwhelmed by potential short-term reward. The medium/long term reward of feeling ok the next morning competes with the short-term reward of drunkenness now. Inter-temporal bargaining can be visualized as a summation of several curves representing hyperbolic discounting of future benefit in a series, which represents a "rule". For example, I could say "I won't drink on Friday nights". If I add these up, the long-term rewards keep on getting collectively bigger, while each short-term reward remains about the same. So that's perhaps how we are able to approximate rational economic behavior. Let me show you a document with some graphs to illustrate: http://www.bbsonline.org/Preprints/Ainslie-0907200 4/Referees/Ainslie.pdf If you are interested you should check out the book Breakdown of Will by George Ainslie, a behavioral economist I think. It is a really rewarding and pretty fast read. I learned about this book by reading some Dennett book (can't remember which one but maybe Consciousness Explained). [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 10/31/2009 10:15 PM by exapted	[Top] [Show Index] [Reply to this post]
	Let me clean up this paragraph: Inter-temporal bargaining can be visualized as a summation of several curves representing hyperbolic discounting of future benefit in a series, which represents a "rule". For example, I could say "I won't drink on Friday nights". If I add these up, the long-term rewards keep on getting collectively bigger, while each short-term reward remains about the same. So that's perhaps how we are able to approximate rational economic behavior. We have two categories of functions. One is the drinking now short-term reward, the other is the not drinking long-term reward on the following day. If we sum the drinking long-term rewards together (a "rule" - by breaking the rule you are likely to lose more than just one instance of the long-term reward due to the nature of habit), we won't be overwhelmed by the short-term reward of drinking now (we don't sum the short-term reward of drinking curves because we don't usually make rules like "I am going to drink every friday night". [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 10/31/2009 11:10 PM by exapted	[Top] [Show Index] [Reply to this post]
	I think we could potentially use ideas like those in the book I mentioned to hard-wire AI for exploitation. Make them dependent on us and train them. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/01/2009 11:36 PM by sensoniq	[Top] [Show Index] [Reply to this post]
	This is interesting exapted, but what about the physical addiction that leads to drinking to merely avoid suffering withdrawel? I suppose that would fall more into the short-term reward modifier, but an AI able to experience painful physical withdrawel (or at least virtual versions of such suffering) may be influenced by a short-term withdrawel-avoidance motivator. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 1:37 AM by exapted	[Top] [Show Index] [Reply to this post]
	Well, in theory, if a robot discounts future benefit exponentially, it might not get addicted. But that depends on whether the benefit of NOT drinking uses the same exponent and/or multiple as the benefit of drinking. If they use the same exponent, then the lines should never cross. But maybe drinking is just very rewarding to a robot, in some subjective sense. So sure, a robot could become addicted I think. I guess you could hard-wire the AI's pico-economics so that nothing could be addictive such that it prevents the robot from deciding to carry out goals that would lead to a certain pre-defined goal. But not all goals can be pre-defined. So I think AIs could suffer from addiction. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 1:39 AM by exapted	[Top] [Show Index] [Reply to this post]
	But with exponential discounting of future benefit, I'm not sure it would really be addictive, because at no point would the AI even want to stop the addiction. An alcoholic with exponential discounting of future benefit where everything gets the same exponent would never regret it, it seems. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 1:43 AM by exapted	[Top] [Show Index] [Reply to this post]
	Well actually that's not right, an AI could regret some faulty reasoning or regret having some drive in the first place, I guess. And anyway AIs don't need to reason like we do. I'm just applying pico-economics to your question and not treating it in the general sense. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 1:42 AM by exapted	[Top] [Show Index] [Reply to this post]
	If you are trying to make a special case out of pleasure from addictive activities, it should be something where future benefit is discounted either hyperbolically or exponentially with a larger exponent/multiple. Then it could suddenly spike above other priorities with time. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 2:16 AM by Phillippe	[Top] [Show Index] [Reply to this post]
	Won't there be pleasure/reward centers in its synthetic brain that are modelled like our own? No, there won't. Why would there be? [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 2:44 AM by exapted	[Top] [Show Index] [Reply to this post]
	An AI might have drives, and I guess not all of them would be pre-defined. It might make sense to allow drives to compete as in a pico-economic model. It would seem rational to apply exponential discounting of future benefit such that rational economic behavior arises. But there may be some benefits to hyperbolic discounting of future benefit. In the case of rational economic behavior, it is possible for one drive to overwhelm the others in a way that, as designers of the AI, we would consider a disadvantage. It could appear to be addiction although as long as all exponents were the same the robot would never "get weak" and change it's mind. But perhaps rational economic behavior would be sub-optimal in some way. It might not be flexible enough. It might be useful to use hyperbolic discounting of future benefit combined with inter-temporal bargaining. However that could lead to the AI "getting weak" when the short-term reward for some unforseen drive overwhelms the other drives - if it is repetitive it is addiction. How could we be sure an AI would be able to avoid such scenarios? [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/02/2009 2:46 AM by exapted	[Top] [Show Index] [Reply to this post]
	Maybe AIs would not have drives, but it is plausible anyway. [Reply] [Parent]

		Re: Will AI suffer from addiction? posted on 11/03/2009 3:36 PM by exapted	[Top] [Show Index] [Reply to this post]
	I just want to correct one thing I said in this thread. I implied that if some AI is discounting future benefit exponentially, and the exponents are different for two different things that the AI might change it's mind. I was forgetting that we're always starting from the same X value, so the exponential functions would never cross each other, no matter what. So if robots discount future benefit exponentially they would never just change their minds unless they had some new information. But they might still "get addicted" if something is just very rewarding to the robot. To the robot it wouldn't really be addiction though. But again maybe hyperbolic discounting of future benefit is generally better than exponential discounting. In that case it would be very difficult to systematically preclude addictive behavior in an AI. [Reply] [Parent]