Hold Off On Proposing Solutions

From pp. 55-56 of Robyn Dawes's Rational Choice in an Uncertain World.  Bolding added.

Norman R. F. Maier noted that when a group faces a problem, the natural tendency of its members is to propose possible solutions as they begin to discuss the problem.  Consequently, the group interaction focuses on the merits and problems of the proposed solutions, people become emotionally attached to the ones they have suggested, and superior solutions are not suggested.  Maier enacted an edict to enhance group problem solving: "Do not propose solutions until the problem has been discussed as thoroughly as possible without suggesting any."  It is easy to show that this edict works in contexts where there are objectively defined good solutions to problems.

Maier devised the following "role playing" experiment to demonstrate his point.  Three employees of differing ability work on an assembly line.  They rotate among three jobs that require different levels of ability, because the most able - who is also the most dominant - is strongly motivated to avoid boredom.  In contrast, the least able worker, aware that he does not perform the more difficult jobs as well as the other two, has agreed to rotation because of the dominance of his able co-worker.  An "efficiency expert" notes that if the most able employee were given the most difficult task and the least able the least difficult, productivity could be improved by 20%, and the expert recommends that the employees stop rotating.  The three employees and the a fourth person designated to play the role of foreman are asked to discuss the expert's recommendation.  Some role-playing groups are given Maier's edict not to discuss solutions until having discussed the problem thoroughly, while others are not.  Those who are not given the edict immediately begin to argue about the importance of productivity versus worker autonomy and the avoidance of boredom.  Groups presented with the edict have a much higher probability of arriving at the solution that the two more able workers rotate, while the least able one sticks to the least demanding job - a solution that yields a 19% increase in productivity.

I have often used this edict with groups I have led - particularly when they face a very tough problem, which is when group members are most apt to propose solutions immediately.  While I have no objective criterion on which to judge the quality of the problem solving of the groups, Maier's edict appears to foster better solutions to problems.

This is so true it's not even funny.  And it gets worse and worse the tougher the problem becomes.  Take Artificial Intelligence, for example.  A surprising number of people I meet seem to know exactly how to build an Artificial General Intelligence, without, say, knowing how to build an optical character recognizer or a collaborative filtering system (much easier problems).  And as for building an AI with a positive impact on the world - a Friendly AI, loosely speaking - why, that problem is so incredibly difficult that an actual majority resolve the whole issue within 15 seconds.  Give me a break.

(Added:  This problem is by no means unique to AI.  Physicists encounter plenty of nonphysicists with their own theories of physics, economists get to hear lots of amazing new theories of economics.  If you're an evolutionary biologist, anyone you meet can instantly solve any open problem in your field, usually by postulating group selection.  Et cetera.)

Maier's advice echoes the principle of the bottom line, that the effectiveness of our decisions is determined only by whatever evidence and processing we did in first arriving at our decisions - after you write the bottom line, it is too late to write more reasons above.  If you make your decision very early on, it will, in fact, be based on very little thought, no matter how many amazing arguments you come up with afterward.

And consider furthermore that We Change Our Minds Less Often Than We Think:  24 people assigned an average 66% probability to the future choice thought more probable, but only 1 in 24 actually chose the option thought less probable.  Once you can guess what your answer will be, you have probably already decided.  If you can guess your answer half a second after hearing the question, then you have half a second in which to be intelligent.  It's not a lot of time.

Traditional Rationality emphasizes falsification - the ability to relinquish an initial opinion when confronted by clear evidence against it.  But once an idea gets into your head, it will probably require way too much evidence to get it out again.  Worse, we don't always have the luxury of overwhelming evidence.

I suspect that a more powerful (and more difficult) method is to hold off on thinking of an answer.  To suspend, draw out, that tiny moment when we can't yet guess what our answer will be; thus giving our intelligence a longer time in which to act.

Even half a minute would be an improvement over half a second.

Comments

[Gray Area]

What circles do you run in Eliezer? I meet a fair number of people who work in AI, (you can say I "work in AI" myself) and so far I can't think of a single person who was sure of a way to build general intelligence. Is this attitude you observe a common one among people who aren't actually doing AI research, but who think about AI?

[Eliezer Yudkowsky]

Oh, I'm not talking about the mainstream AI field. Most of them know better. I mean, say, a random middle or upper-class individual in Silicon Valley, or a random user on an IRC channel.

However, the rule about instantly solving Friendly AI may apply even within the AI field, since it's a more difficult problem.

[Constant]

It's obvious how to build AI. You just add complexity. AIs need complexity. :-)

[Richard Hollerith]

And a randomness-adder :)

[Eddieosh]

I've just finished a 3-day training course on TRIZ (http://en.wikipedia.org/wiki/TRIZ) a problem solving technique, one of the recurring themes throughout the course was what to do about all the solutions that come out even before you've figured out what the true problem is you're trying to solve. The advice was to write the solutions down (rather than be diverted by them or try to bat them away), use them to help examine the problem a bit more and then carry on until you have enough information to make useful judgements about all the solutions you've generated; this was very helpful advice. You need to have a sound way of formulating and exploring the problem space, as well as generating solutions, otherwise you'll become too distracted by all the great solutions your brain is generating.

[logicnazi]

I just want to remark that it is far from obvious on apriori grounds that there is no elegant general AI algorithm that will solve all the other problems quite nicely. We've only learned this by the continued failure to find such an algorithm or anything like it by the AI community and the continued small successes of more specific less elegant approaches.

[Rick Smith]

AI's need Emergence too. Make sure to add some of that to the soup ;^)

[Alan Crowe]

X3J13, the ANSI committee that standarised Common Lisp, had many problems to solve. Kent Pitman credits Larry Masinter with imposing the disciple of seperating problem descriptions from proposed solutions and gives insights into what that meant in practise in a post to comp.lang.lisp

http://tinyurl.com/2hppgs

The general interest lies in that fact that the X3J13 Issues were all written up and are available on line.

http://www.lispworks.com/documentation/HyperSpec/Front/X3J13Iss.htm

or

http://www.lisp.org/HyperSpec/FrontMatter/X3J13-Issues.html

so if you wish to study how this works there is a resource you can analyse.

I should confess that my interest has been in content not process. I have been reading these issues to learn Common Lisp. Are these pages really a useful resource for scholars wishing to study the separation of problem descriptions from proposed solutions? I don't know.

[Tiedemies]

I think this argument is flawed with respect to the more technology-oriented questions. Most people do not seriously claim to solve AI problems. What most people (like myself) who are slightly educated in the field (I did an undergrad minor in AI, just very simple stuff) will do is they will suggest an approach that they would try if they had to start working on it. Technical questions also usually yield to evidence very quickly whenever it matters, i.e., when someone would start burning money on an implementation. That is not to say some time and resources are not to be saved by using the maxim outlined here.

OTOH, the part about economists is valid, since most people have very strong ideas (usually wrong ones) about what will work, e.g., as a policy. But then again, most people have no way of wasting (other peoples') resources based on these faulty ideas.

No, wait...

[4σ]

The latest of a number of really good posts from you that directly address the concern of this blog. You seem to be really starting to "grok" the terrifying reality of just how biased we are by the very nature of our thought processes, and coming up with good and useful steps to reduce those biases. Nicely done.

[michael vassar]

This post makes me wonder how much time passed for Eliezer between concluding that a technological singularity was a probable part of the future and deciding that creating an AGI was the best response, and likewise how much time passed between concluding that AGI Friendlyness would be a difficult problem and concluding that working on a theory of AGI Friendlyness was the best response.

[Richard Hollerith]

Eliezer, I get the impression that your recent blog entries will make me a better rationalist or if not that a better inventor of software, organizational innovations and social arrangements that will help people become better rationalists.

Good stuff, I say.

[anonymous poster]

A surprising number of people I meet seem to know exactly how to build an Artificial General Intelligence, without, say, knowing how to play the guitar or juggle (much easier problems).

[Constant]

My AI will play the guitar and juggle so I won't have to.

[Constant]

This advice seems the opposite of, "avoid analysis paralysis." These may be bounding two extremes, neither of which is healthy. Or I may simply be wrong about the relationship.

[Eliezer Yudkowsky]

Playing the guitar has human-aesthetic components so it's a subproblem of Friendly AI, not just AGI. Building an AI that juggles is a valid challenge. As for trying to do it yourself, that quite misses the point. A mathematician may not be able to do high-speed mental arithmetic, but ought to know how to build a calculator.

[Luis Enrique]

I remember reading something much like this in I am right and you are wrong by Edward de Bono, who as I recall wrote that we should try to hold on to the "I haven't made my mind up" state much longer than we do, and be prepared to say "I don't know" much more often than we do (I think he even proposed a new word we could use to answer questions with that meant we don't have a reason to think either way yet). This was about 15 years ago so I've probably mis-remembered.

I was a philosophy undergrad at the time, and when I asked my tutors about de Bono, they told me he was a vacuous 'self-help' nitwit I should ignore.

[GreedyAlgorithm]

"My Ap distribution is rather flat."

Hm, MADIRF? :)

[Doug S.]

Completely useless methods for building a general intelligence:

Method 1:
Put some bacteria on a lifeless planet with liquid water. Wait until one evolves.

Method 2:
Find a fertile human of each gender and induce them to mate. Wait nine months.

[Douglas Knight]

Luis Enrique,
See above about "We Change Our Minds Less Often Than We Think"; my interpretation is that the people are trying to believe that they haven't made up their minds, but they are wrong. That is, they seem to be implementing the (first) advice you mention. Maybe one can come up with more practical advice, but these are very difficult problems to fix, even if you understand the errors. On the other hand, the main part of the post is about a successful intervention.

[Rolf Nelson]

Constant, regarding "analysis paralysis," keep in mind there are often two separate questions:

1. How much time should I spend thinking about X?

2. Given I'm allocating T time to think about X, how should I divide up T among different thought subtasks?

Analysis Paralysis would generally be a problem with (1).

The current blog post applies more to (2). In the Maier example, the participants presumably know they have a sizable chunk of time blocked out, and the experimental group presumably gets better results not by spending more time overall, but because they reserved a good chunk of T to spend learning the problem, *without* committing right away to a solution.