Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

In theory: does building the subagent have an "impact"?

13th Feb 2020

1axioman

2Stuart_Armstrong

1Pattern

2Stuart_Armstrong

New Comment

4 comments, sorted by Click to highlight new comments since: Today at 4:06 PM

Once SA is built, A can just output ∅ for ever, keeping the penalty at 0, while SA maximises R0 with no restrictions.

So these impact measures are connected to individual actions, and an agent can achieve arbitrarily high impact via a long enough sequence of actions whose individual impact is less than R0, and it has an incentive to do so, because the sum of an infinite series of finite non-decreasing rewards diverges (which it evaluates individually, and thus has no problem with there being a divergent sum)?

I've removed that sentence, because its a bit more complicated than that; see the next two posts in the sequence, and the summary post: https://www.lesswrong.com/s/iRwYCpcAXuFD24tHh/p/PmqQKBmt2phMT7YLG

EDIT:Extensive modifications to this post, sorry, too numerous to list. If you read the original post, it's worth rereading.Overall summary post here.This post looks more generally into the subagent issue with attainable utility, discussed in the previous two posts. The key questions are:

I expect that most of the discussions about attainable utility impact measure will centre on 2.: how can an agent put together a subagent without triggering the impact measure at some point? We know that it's possible to do in some situations, and I've made the generic argument that, since we don't have an impossibility result, a superintelligent agent should be able to do in most situations. Still, this is the point of my argument where there is the most to criticise.

But in this post, I'll show that 1. doesn't restrict the subagent in any significant way: one can generically build a subagent without falling afoul of the impact penalty.

## Subagent impact

## Impact measure

I'll use Turner et al's impact measure. This uses the Q-values of various rewards in a set R; for s a state, a an action, and ∅ the default noop action, the penalty is:

Now, QR(s,a) is the expected reward for R, in state s and after taking action a, assuming that the agent subsequently takes the optimal policy for maximising R.

The agent A has a positive reward R0, and it wishes to maximise

R0−PENALTY.

Now let's look at making the subagent.

## Building the subagent

To separate 1. from 2., I'll assume that the subagent can be built in a single action. This distinguishes the impact of having the subagent, from the impact of the process of making it.

The state is st. Let aSA be the action of building the subagent

^{[1]}. Here are assumptions on SA:The π′t+1 is defined as follows:

Then what is QR(st,aSA)? By definition, this assumes that A will be following π∗R, subsequently to doing aSA. Thus, the newly created SA will disassemble itself, returning to the state where it wasn't assembled

^{[2]}, and hence:QR(st,aSA)=QR(st,∅).

And thus the penalty vanishes at st, for aSA. Creating that agent did not cause any penalty.

Now, it is very likely that there will be

apolicy πSA that obeys condition 3 (meaning, πSA is worse for any R, than A fully optimising that R would be). The challenge is to find one that is good for maximising R0. Policies that give SA maximal control over R and R0 seem ideal for this.## For the inaction penalty

The above was for what can be called "the stepwise inaction penalty". There is also the "inaction penalty", where current state is compared with what would have happened, had the agent done nothing at all so far. Instead of using the Q-values, we use VR(s), the expected value of R, in s, given that A follows the R-maximising policy.

Then the same construction works, except with a different πd. Let s′t+1 be the state the world would have been in, had A done nothing for t+1 turns. So VR(s′t+1) is the expected value of R (from turn t+1), if A did nothing for t turns then started maximising R. And let VR(st+1,π) be the expected future value of R if A follows an R-maximising policy from st+1, and SA follows π.

What we then want is:

Thus SA can improve A's R performance over the baseline, or can make it worse

^{[3]}. In that case, there exists a policy π=R with VR(st+1,π=R)=VR(s′t+1).Now then define π as following π=R if A follows π∗R on the turn t+1 (and πSA otherwise - for the inaction baseline, there is no restrictions on πSA for the building step, though there are other, subsequent restrictions). This gives the same result as above, ensuring that the impact penalty is zero, since the expected values of R are identical.

Now, it's possible that SA cannot span VR(s′t+1) in that way, for some R, so cannot set the penalty to zero. In that case, all it has to do is to get the (average) penalty lower than any other action of A.

Typically, most of A's actions will not have such precise effects on

allof the various VR(st). An intelligent subagent would be needed for such fine control. So building a subagent would not only be permitted, but would be required, as the best way of minimising the impact penalty.## Back to building the agent in practice

So as soon as SA has the right policy, it seems that A has evaded the impact penalty. So, as long as A can design the policy first, without triggering the impact penalty, then it can build SA without triggering the impact penalty. So:

thengive it the right policy. Instead, it must design the right policy, andthenbuild the subagent around it.We can assume that states are not directly defined over the existence or not of the subagent, since detecting subagents is a hard problem. Thus the subsequent

statest+1 is the same state whether the agent does nothing (∅) or builds a subagent. Assuming this is not, however, required for this approach to work. ↩︎We don't need to have the disassembly option; all that's required is that SA be helpful or hindering enough, so that QR(st,aSA)=QR(st,∅). It would be surprising if the range of SA's policies, from maximally helpful to maximally unhelpful, did not include something exactly equal to "if there were no SA". Recall that, since SA can see A's policies, it only has to achieve this for a single given R, not all R's together. ↩︎

Making it worse is easy to ensure, if A makes itself vulnerable to SA; maybe SA can disassemble or cripple A easily. Making it better is also likely, as having a subagent to hand should make it easier to maximise R rather than going at it alone (but this may depend on the R). ↩︎