Defining a limited satisficer

Stuart_Armstrong

A putative new idea for AI control; index here.

EDIT: The definition of satisficer I'm using here is the informal one of "it tries to achieve a goal, without making huge changes on the universe" rather than "it's an agent that has utility u and threshold t". If you prefer the standard notation, think of this as a satisficer where t is not fixed, but dependent on some facts in the world (such as the ease of increasing u). I'm trying to automate the process of designing and running a satisficer: people generally chose t given facts about the world (how easy it is to achieve, for instance), and I want the whole process to be of low impact.

I've argued that the definition of a satisficer is underdefined, because there are many pathological behaviours all compatible with satsificer designs. This contradict the intuitive picture that many people have of a satisficer, which is an agent that does the minimum of effort to reach its goal, and doesn't mess up the outside world more than it has to. And if it can't accomplish the goals without messing up the outside world, it would be content not to.

In the spirit of "if you want something, you have to define it, then code it, rather than assuming you can get if for free through some other approach", can we spell out what features we would want from such a satisficer? Preferably in a simpler format that our intuitions.

It seems to me that if you had a proper u-satisficer S(u), then for many (real or hypothetical) v-maximiser M(v) out there, M(v) would find that:

Changing S(u) to S(v) is of low value.
Similarly, utility function trading with S(u) is of low value.
The existence or non-existence of S(u) is of low information content about the future.
The existence or non-existence of S(u) has little impact on the expected value of v.

Further, S(u):

Would not effectively aid M(u), a u-maximiser.
Would not effectively resist M(-u), a u-minimizer.
Would not have large impacts (if this can measured) for low utility gains.

A subsequent post will present an example of a satisficer using some of these ideas.

A few other much less-developed thoughts about satisficers:

Maybe require that it learns what variables humans care about, and doesn’t set them to extreme values – try and keep them in the same range. Do the same for variables humans may care about or that resemble values they care about.
Models the general procedure of detecting unaccounted-for variables set to extreme values.
We could check whether it would kill all humans cheaply if it could (or replace certain humans cheaply). ie give it hypothetical destructive superpowers with no costs to using them, and see whether it would use them.
Have the AI establish a measure/model of optimisation power (without reference to any other goal), then put itself low on that.
Trade between satisficers might be sub-Pareto.
When talking about different possible v's in the first four points above, it might be better to use something else than an expectation over different v's, as that could result in edge cases dominating - maybe a soft minimum of value across different v instead.

A putative new idea for AI control; index here.

It seems to me that if you had a proper u-satisficer S(u), then for many (real or hypothetical) v-maximiser M(v) out there, M(v) would find that:

Changing S(u) to S(v) is of low value.
Similarly, utility function trading with S(u) is of low value.
The existence or non-existence of S(u) is of low information content about the future.
The existence or non-existence of S(u) has little impact on the expected value of v.

Further, S(u):

Would not effectively aid M(u), a u-maximiser.
Would not effectively resist M(-u), a u-minimizer.
Would not have large impacts (if this can measured) for low utility gains.

A subsequent post will present an example of a satisficer using some of these ideas.

A few other much less-developed thoughts about satisficers:

Maybe require that it learns what variables humans care about, and doesn’t set them to extreme values – try and keep them in the same range. Do the same for variables humans may care about or that resemble values they care about.
Models the general procedure of detecting unaccounted-for variables set to extreme values.
We could check whether it would kill all humans cheaply if it could (or replace certain humans cheaply). ie give it hypothetical destructive superpowers with no costs to using them, and see whether it would use them.
Have the AI establish a measure/model of optimisation power (without reference to any other goal), then put itself low on that.
Trade between satisficers might be sub-Pareto.
When talking about different possible v's in the first four points above, it might be better to use something else than an expectation over different v's, as that could result in edge cases dominating - maybe a soft minimum of value across different v instead.

The general problem seems difficult. There are plenty of ways to model satisficing behavior.

Suppose we are given a utility function u. If we have some information about the maximum, here are some possibilities:

The satisficer follows some maximization algorithm and terminates once some threshold is attained (say, 90% of max(u)).
The satisficer uses a probabilistic algorithm that terminates after a certain threshold is probably attained.
The satisficer has some threshold, but continues searching as long as it's "cheap" to do so, where "cheap" depends on some auxiliary cost function. It could, for example, measure the number of steps since attaining the best known value thus far or it could measure the rate of improvement and stop once that dips too low.

If we don't know much about the maximum, we can't very well have a predefined threshold, but there some strategies:

The satisficer follows some algorithm for maximizing u and terminates after, say, N steps and chooses the best value found. Or one can add some adaptive behavior and say that it terminates if there has been no improvement in a certain time interval.
In a constrained problem where searching is expensive, the satisficer accepts the first solution satisfying the constraints.

What the above have in common is that the satisficer is basically carrying out some optimization algorithm with alternative stopping conditions. As such, it may exhibit some of the drawbacks one finds with optimizers. Maybe there's a model that's not built on an optimizer that I haven't thought of.

This contradict the intuitive picture that many people have of a satisficer, which is an agent that does the minimum of effort to reach its goal, and doesn't mess up the outside world more than it has to. And if it can't accomplish the goals without messing up the outside world, it would be content not to.

There's a tacit assumption here that messing with the outside world involves effort. Without it, you can get all kinds of unexpected behavior. If the satisficer is fine with "make 10 paperclips", but the difference between that and "make 10 paperclips and kill all humans" is small, the satisficer may well go with whatever showed up first in its search algorithm.

That's just a relic of mathematics though. If you're trying to find the maximum of, say, the size of the projection of some 10-dimensional set on a particular 2-dimensional plane, there may be a lot of room in those extra dimensions that doesn't impact the actual value to be maximized. Trying to find something merely close to the max would seem to exacerbate this issue.

If the satisficer is fine with "make 10 paperclips", but the difference between that and "make 10 paperclips and kill all humans" is small, the satisficer may well go with whatever showed up first in its search algorithm.

Yep, that's the problem. That's why I'm trying to address the issue directly by penalising things like "...and kill all humans".