A putative new idea for AI control; index here.

EDIT: I feel this post is unclear, and will need to be redone again soon.

This post attempts to use the ideas developed about natural categories in order to get high impact from reduced impact AIs.

 

Extending niceness/reduced impact

I recently presented the problem of extending AI "niceness" given some fact X, to niceness given ¬X, choosing X to be something pretty significant but not overwhelmingly so - the death of a president. By assumption we had a successfully programmed niceness, but no good definition (this was meant to be "reduced impact" in a slight disguise).

This problem turned out to be much harder than expected. It seems that the only way to do so is to require the AI to define values dependent on a set of various (boolean) random variables Z_j that did not include X/¬X. Then as long as the random variables represented natural categories, given X, the niceness should extend.

What did we mean by natural categories? Informally, it means that X should not appear in the definitions of these random variables. For instance, nuclear war is a natural category; "nuclear war XOR X" is not. Actually defining this was quite subtle; diverting through the grue and bleen problem, it seems that we had to define how we update X and the Z_j given the evidence we expected to find. This was put in equation as picking Z_j's that minimize

• Variance{log[ P(X∧Z|E)*P(¬X∧¬Z|E) / P(X∧¬Z|E)*P(¬X∧Z|E) ]} 

where E is the random variable denoting the evidence we expected to find. Note that if we interchange X and ¬X, the ratio inverts, the log changes sign - but this makes no difference to the variance. So we can equally well talk about extending niceness given X to ¬X, or niceness given ¬X to X.

 

Perfect and imperfect extensions

The above definition would work for an "perfectly nice AI". That could be an AI that would be nice, given any combination of estimates of X and Z_j. In practice, because we can't consider every edge case, we would only have an "expectedly nice AI". That means that the AI can fail to be nice in certain unusual and unlikely edge cases, in certain strange set of values of Z_j that almost never come up...

...or at least, that almost never come up, given X. Since the "expected niceness" was calibrated given X, the such an expectedly nice AI may fail to be nice if ¬X results in a substantial change in the probability of the Z_j (see the second failure mode in this post; some of the Z_j may be so tightly coupled to the value of X that an expected niceness AI considers them fixed, and this results in problems if ¬X happens and their values change).

One way of fixing this is to require that the "swing" of the Z_j be small upon changing X to ¬X or vice versa. Something like, for all values of {a_j}, the ratio P({Z_j=a_j} | X) / P({Z_j=a_j} | ¬X) is contained between 100 and 1/100. This means that a reasonably good "expected niceness" calibrated on the Z_j will transfer from X to ¬X (though the error may grow). This approach has some other advantages, as we'll see in the next section.

Of course, problems arise if the Z_j are defined by smashing together events relatively independent of X with some that are very strongly dependent on X, into the same variable. Thus we'll further require that Z_j cannot be decomposed into natural subvariables whose values swing more under changes from X to ¬X.

 

High impact from reduced impact

With this, we are finally in a position to define a reduced impact AI design that can have a large impact. Imagine that an asteroid is coming to hit the Earth. Two AIs are charged with targeting a laser at it. The laser isn't perfectly effective, and the asteroid isn't perfectly certain to hit the Earth anyway: let's say aiming the laser correctly will reduce the odds of impact by a factor of 10 or so (well within the probability swing restrictions of the previous section).

In the original formulation of this, the AIs were tasked with outputting the x and y coordinates to aim the laser. Here, instead, they'll be tasked with constructing the motivation of a robotic subagent that will move around for a while, see things and interact, before physically aiming the laser in the right direction. This task is harder.

Then the top level AI charged with aiming the y coordinates is given the following task:

• It is motivated to cause the correct y-coordinate to be outputted.
• If the x-coordinate AI is not turned on (call this event ¬X), it is strongly motivated to have reduced impact.
• It must phrase the robotic subagent's utility function in terms of the Z_j that obey all the properties above.
• The evidence E considered for the equation governing the naturalness of Z_j is considered (for example) to be a random sampling of space at the time both AIs are turned on. Most importantly, this is defined to be a theoretical process that the AI cannot affect physically.

So what will the AI do? It will produce a robot that will aim to aim the y-coordinate of the laser correctly, given ¬X, and not expand dangerously. However, we've setup things so that X is virtually certain, so the robot will see another, x-aiming robot, wander around. The chances of this happening, given ¬X, are minuscule, so the swing of the probability from ¬X to X, is huge. So, by the restrictions we've put on the Z_j, the robot will not be motivated to change its behaviour just because its sees the other robot.

And if the laser is fired, the impact on humanity will, conversely, not be that large, so well within the parameters of the Z_j. Thus the "general reduced impact" portion of the robot's motivation will be preserved.

And, of course, the x-robot will have the mirror image motivations, so they will both aim correctly, it seems. This should evade the problem described here.

 

Failure modes

There are two ways the approach could fail: one bad way, and two graceful ways. The bad way is if the "no decomposition" requirements (one of them is presented above, the other is in this post) fail. In this case, the niceness or reduced impact are not guaranteed to be extended to ¬X, and the AI's behaviour can become undefined. Note that this is is not bad optimisation pressure, just an absence of optimisation pressure - but that could still be potentially dangerous.

The first graceful failure happens if the AI is not able to create a reduced impact sub-agent given the restrictions on the Z_j above. In that case, the AI does nothing. The second graceful failure happens if the AI evades our attempts to increase its impact, given ¬X. In that case, it simply becomes a reduced impact AI that does little. Not ideal, but not deadly.

 

Overall status: I'm not sure the idea is sound, at least not yet. Critiques welcome.

A putative new idea for AI control; index here.

In a previous post, I looked at unnatural concepts such as grue (green if X was true, blue if it was false) and bleen. This was to enable one to construct the natural categories that extend AI behaviour, something that seemed surprisingly difficult to do.

The basic idea discussed in the grue post was that the naturalness of grue and bleen seemed dependent on features of our universe - mostly, that it was easy to tell whether an object was "currently green" without knowing what time it was, but we could not know whether the object was "currently grue" without knowing the time.

So the naturalness of the category depended on the type of evidence we expected to find. Furthermore, it seemed easier to discuss whether a category is natural "given X", rather than whether that category is natural in general. However, we know the relevant X in the AI problems considered so far, so this is not a problem.

 

Natural category, probability flows

Fix a boolean random variable X, and assume we want to check whether the boolean random variable Z is a natural category, given X.

If Z is natural (for instance, it could be the colour of an object, while X might be the brightness), then we expect to uncover two types of evidence:

• those that change our estimate of X; this causes probability to "flow" as follows (or in the opposite directions):

• ...and those that change our estimate of Z:

Or we might discover something that changes our estimates of X and Z simultaneously. If the probability flows to X and and Z in the same proportions, we might get:

What is an example of an unnatural category? Well, if Z is some sort of grue/bleen-like object given X, then we can have Z = X XOR Z', for Z' actually a natural category. This sets up the following probability flows, which we would not want to see:

More generally, Z might be constructed so that X∧Z, X∧¬Z, ¬X∧Z and ¬X∧¬Z are completely distinct categories; in that case, there are more forbidden probability flows:

and

In fact, there are only really three "linearly independent" probability flows, as we shall see.

 

Less pictures, more math

Let's represent the four possible state of affairs by four weights (not probabilities):

Since everything is easier when it's linear, let's set w₁₁ = log(P(X∧Z)) and similarly for the other weights (we neglect cases where some events have zero probability). Weights are correspond to the same probabilities iff you get from one set to another by multiplying by a strictly positive number. For logarithms, this corresponds to adding the same constant to all the log-weights. So we can normalise our log-weights (select a single set of representative log-weights for each possible probability sets) by choosing the w such that

w₁₁ + w₁₂ + w₂₁ + w₂₂ = 0.

Thus the probability "flows" correspond to adding together two such normalised 2x2 matrices, one for the prior and one for the update. Composing two flows means adding two change matrices to the prior.

Four variables, one constraint: the set of possible log-weights is three dimensional. We know we have two allowable probability flows, given naturalness: those caused by changes to P(X), independent of P(Z), and vice versa. Thus we are looking for a single extra constraint to keep Z natural given X.

A little thought reveals that we want to keep constant the quantity:

w₁₁ + w₂₂ - w₁₂ - w₂₁.

This preserves all the allowed probability flows and rules out all the forbidden ones. Translating this back to a the general case, let "e" be the evidence we find. Then if Z is a natural category given X and the evidence e, the following quantity is the same for all possible values of e:

log[P(X∧Z|e)*P(¬X∧¬Z|e) / P(X∧¬Z|e)*P(¬X∧Z|e)].

If E is a random variable representing the possible values of e, this means that we want

log[P(X∧Z|E)*P(¬X∧¬Z|E) / P(X∧¬Z|E)*P(¬X∧Z|E)]

to be constant, or, equivalently, seeing the posterior probabilities as random variables dependent on E:

• Variance{log[ P(X∧Z|E)*P(¬X∧¬Z|E) / P(X∧¬Z|E)*P(¬X∧Z|E) ]} = 0.

Call that variance the XE-naturalness measure. If it is zero, then Z defines a XE-natural category. Note that this does not imply that Z and X are independent, or independent conditional on E. Just that they are, in some sense, "equally (in)dependent whatever E is".

 

Almost natural category

The advantage of that last formulation becomes visible when we consider that the evidence which we uncover is not, in the real world, going to perfectly mark Z as natural, given X. To return to the grue example, though most evidence we uncover about an object is going to be the colour or the time rather than some weird combination, there is going to be somebidy who will right things like "either the object is green, and the sun has not yet set in the west; or instead perchance, those two statements are both alike in falsity". Upon reading that evidence, if we believe it in the slightest, the variance can no longer be zero.

Thus we cannot expect that the above XE-naturalness be perfectly zero, but we can demand that it be low. How low? There seems no principled way of deciding this, but we can make one attempt: that we cannot lower it be decomposing Z.

What do we mean by that? Well, assume that Z is a natural category, given X and the expected evidence, but Z' is not. Then we can define a new category boolean Y to be Z with high probability, and Z' otherwise. This will still have low XE-naturalness measure (as Z does) but is obviously not ideal.

Reversing this idea, we say Z defines a "XE-almost natural category" if there is no "more XE-natural" category that extends X∧Z (and the other for conjunctions). Technically, if

X∧Z = X∧Y,

Then Y must have equal or greater XE-naturalness measure to Z. And similarly for X∧¬Z, ¬X∧Z, and ¬X∧¬Z.

Note: I am somewhat unsure about this last definition; the concept I want to capture is clear (Z is not the combination of more XE-natural subvariables), but I'm not certain the definition does it.

 

Beyond boolean

What if Z takes n values, rather than being a boolean? This can be treated simply.

If we set the w_jk to be log-weights as before, there are 2n free variables. The normalisation constraint is that they all sum to a constant. The "permissible" probability flows are given by flows from X to ¬X (adding a constant to the first column, subtracting it from the second) and pure changes in Z (adding constants to various rows, summing to 0). There are 1+ (n-1) linearly independent ways of doing this.

Therefore we are looking for 2n-1 -(1+(n-1))=n-1 independent constraints to forbid non-natural updating of X and Z. One basis set for these constraints could be to keep constant the values of

w_j1 + w_(j+1)2 - w_j2 - w_(j+1)1,

where j ranges between 1 and n-1.

This translates to variance constraints of the type:

• Variance{log[ P(X∧{Z=j}|E)*P(¬X∧{Z=j+1}|E) / P(X∧{Z=j+1}|E)*P(¬X∧{Z=j}|E) ]} = 0.

But those are n different possible variances. What is the best global measure of XE-naturalness? It seems it could simply be

• Max_jk Variance{log[ P(X∧{Z=j}|E)*P(¬X∧{Z=k}|E) / P(X∧{Z=k}|E)*P(¬X∧{Z=j}|E) ]} = 0.

If this quantity is zero, it naturally sends all variances to zero, and, when not zero, is a good candidate for the degree of XE-naturalness of Z.

The extension to the case where X takes multiple values is straightforward:

• Max_jklm Variance{log[ P({X=l}∧{Z=j}|E)*P({X=m}∧{Z=k}|E) / P({X=l}∧{Z=k}|E)*P({X=m}∧{Z=j}|E) ]} = 0.

Note: if ever we need to compare the XE-naturalness of random variables taking different numbers of values, it may become necessary to divide these quantities by the number of variables involved, or maybe substitute a more complicated expression that contains all the different possible variances, rather than simply the maximum.

 

And in practice?

In the next post, I'll look at using this in practice for an AI, to evade presidential deaths and deflect asteroids.