= 27d567bc2d48d9f40e9ccc20ea370b15)
orthonormal comments on Rationality Quotes: November 2010 - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (354)
Perplexed, have you come across the decision theory posts here yet? You'll find them pretty interesting, I think.
LW Wiki for the Prisoner's Dilemma
LW Wiki for timeless decision theory (start with the posts- Eliezer's PDF is very long and spends more time justifying than explaining).
Essentially, this may be beyond the level of humans to implement, but there are decision theories for an AI which do strictly better than the usual causal decision theory, without being exploitable. Two of these would cooperate with each other on the PD, given a chance to communicate beforehand.
Yes, I have read them, and commented on them. Negatively, for the most part. If any of these ideas are ever published in the peer reviewed literature, I will be both surprised and eager to read more.
I think that you may have been misled by marketing hype. Even the proponents of those theories admit that they do not do strictly better (or at least as good) on all problems. They do better on some problems, and worse on others. Furthermore, sharing source code only provides a guarantee that the observed source is current if that source code cannot be changed. In other words, an AI that uses this technique to achieve commitment has also forsaken (at least temporarily) the option of learning from experience.
I am intrigued by the analogy between these acausal decision theories and the analysis of Hamilton's rule in evolutionary biology. Nevertheless, I am completely mystified as to the motivation that the SIAI has for pursuing these topics. If the objective is to get two AIs to cooperate with each other there are a plethora of ways to do that already well known in the game theory canon. An exchange of hostages, for example, is one obvious way to achieve mutual enforceable commitment. Why is there this fascination with the bizarre here? Why so little reference to the existing literature?
So far as I understand the situation, the SIAI is working on decision theory because they want to be able to create an AI that can be guaranteed not to modify its own decision function.
There are circumstances where CDT agents will self-modify to use a different decision theory (e.g. Parfit's Hitchhiker). If this happens (they believe), it will present a risk of goal-distortion, which is unFriendly.
Put another way: the objective isn't to get two AIs to cooperate, the objective is to make it so that an AI won't need to alter its decision function in order to cooperate with another AI. (Or any other theoretical bargaining partner.)
Does that make any sense? As a disclaimer, I definitely do not understand the issues here as well as the SIAI folks working on them.
I don't think that's quite right- a sufficiently smart Friendly CDT agent could self-modify into a TDT (or higher decision theory) agent without compromising Friendliness (albeit with the ugly hack of remaining CDT with respect to consequences that happened causally before the change).
As far as I understand SIAI, the idea is that decision theory is the basis of their proposed AI architecture, and they think it's more promising than other AGI approaches and better suited to Friendliness content.
That sounds intriguing also. Again, a reference to something written by someone who understands it better might be helpful so as to make some sense of it.
Maybe it would be helpful to you to think of self-modifications and alternative decision theories as unrestricted precommitment. If you had the ability to irrevocably precommit to following any decision rule in the future, which rule would you choose? Surely it wouldn't be pure CDT, because you can tractably identify situations where CDT loses.
"Tractably" is a word that I find a bit unexpected in this context. What do you mean by it?
"Situations where CDT loses." Are we talking about real-world-ish situations here? Situations in which causality applies? Situations in which the agents are free rather than being agents whose decisions have already been made for them by a programmer at some time in the past? What kind of situations do you have in mind?
And what do you mean by "loses"? Loses to who or what? Loses to agents that can foresee their opponent's plays? Agents that have access to information channels not available to the CDT agent? Just what information channels are allowed? Why those, and not others?
ETA: And that "Surely it wouldn't be CDT ... because you can identify ..." construction simply begs for completion with "Surely it would be <my candidate> ... because you can't identify ...". Do you have a candidate? Do you have a proof of "you can't identify situations where it loses". If not, what grounds do you have for criticizing?
CDT still loses to TDT in Newcomb's problem if Omega has can predict your actions with better than 50.05% chances. You can't get out of this by claiming that Omega has access to unrealistic information channels, because these chances seem fairly realistic to me.
Free from what? Causality? This sounds distressingly like you are relying on some notion of "free will".
(Apologies if I'm misreading you.)
I am relying on a notion of free will.
I understand that every normative decision theory adopts the assumption (convenient fiction if you prefer) that the agent being advised is acting of "his own free will". Otherwise, why bother advising?
Being a compatibilist, as I understand Holy Scripture (i.e. The Sequences) instructs me to be, I see no incompatibility between this "fiction" of free will and the similar fiction of determinism. They model reality at different levels.
For certain purposes, it is convenient to model myself and other "free agents" as totally free in our decisions, but not completely free in carrying out those decisions. For example, my free will ego may decide to quit smoking, but my determined id has some probability of overruling that decision.
Why the distinction between agents which are free and agents which have had their decisions made for them by a programmer, then? Are you talking about cases in which specific circumstances have hard-coded behavioral responses? Every decision every agent makes is ultimately made for it by the agent's programmer; I suppose I'm wondering where you draw the line.
As a side note, I feel very uncomfortable seeing the sequences referred to as inviolable scripture, even in jest. In my head, it just screams "oh my god how could anyone ever be doing it this wrong arghhhhhh."
I'm still trying to figure out what I think of that reaction, and do not mention it as a criticism. I think.
"Totally free" sounds like too free. You're not free to actually decide at time T to "decide X at time T+1" and then actually decide Y at time T+1, since that is against the laws of physics.
It's my understanding that what goes through your head when you actually decide X at time T+1 is (approximately) what we call TDT. Or you can stick to CDT and not be able to make decisions for your future self.
Situations where CDT loses are precisely those situations where credible precommitment helps you, and inability to credibly precommit hurts you. There's no shortage of those in game theory.
Ok, those are indeed a reasonable class of decisions to consider. Now, you say that CDT loses. Ok, loses to what? And presumably you don't mean loses to opponents of your preferred decision theory. You mean loses in the sense of doing less well in the same situation. Now, presumably that means that both CDT and your candidate are playing against the same game opponent, right?
I think you see where I am going here, though I can spell it out if you wish. In claiming the superiority of the other decision theory you are changing the game in an unfair way by opening a communication channel that didn't exist in the original game statement and which CDT has no way to make use of.
Well, yeah, kind of, that's one way to look at it. Reformulate the question like this: what would CDT do if that communication channel were available? What general precommitment for future situations would CDT adopt and publish? That's the question TDT people are trying to solve.
I'm going to have to refer you to Eliezer's TDT document for that. (If you're OK with starting in medias res, the first mention of this is on pages 22-23, though there it's just specialized to Newcomb's Dilemmas; see pages 50-52 for an example of the limits of this hack. Elsewhere he's argued for the more general nature of the hack.)
Ok thanks.
I'm coming to realize just how much of this stuff derives from Eliezer's insistance on reflective consistency of a decision theory. Given any decision theory, Eliezer will find an Omega to overthrow it.
But doesn't a diagonal argument show that no decision theory can be reflectively consistent over all test data presented by a malicious Omega? Just as there is no enumeration of the reals, isn't there a game which can make any specified rational agent regret its rationality? Omega holds all the cards. He can always make you regret your choice of decision theory.
No. We can ensure that no such problem exists if we assume that (1) only the output decisions are used, not any internals; and (2) every decision is made with access to the full problem statement.
I'm not entirely sure what "every decision is made with full access to the problem statement means", but I can't see how it can possibly get around the diagonalisation argument. Basically, Omega just says "I simulated your decision on problem A, on which your algorithm outputs something different from algorithm X, and give you a shiny black ferrari iff you made the same decision as algorithm X"
As cousin_it pointed out last time I brought this up, Caspian made this argument in response to the very first post on the Counterfactual Mugging. I've yet to see anyone point out a flaw in it as an existence proof.
As far as I can see the only premise needed for this diagonalisation to work is that your decision theory doesn't agree with algorithm X on all possible decisions, so just make algorithm X "whatever happens, recite the Bible backwards 17 times".
In that case, your answer to problem A is being used in a context other than problem A. That other context is the real problem statement, and you didn't have it when you chose your answer to A, so it violates the assumption.
Yeah, that definitely violates the "every decision is made with full access to the problem statement" condition. The outcome depends on your decision on problem A, but when making your decision on problem A you have no knowledge that your decision will also be used for this purpose.
This is a no-choice scenario. If you say that the Bible-reciter is the one that will "win" here, you are using the verb "to win" with a different meaning from the one used when we say that a particular agent "wins" by making the choice that leads to the best outcome.
With the strong disclaimer that I have no background in decision theory beyond casually reading LW...
I don't think so. The point of simulation (Omega) problems, to me, doesn't seem to be to judo your intelligence against yourself; rather, it is to "throw your DT off the scent", building weird connections between events (weird, but still vaguely possible, at least for AIs), that a particular DT isn't capable of spotting and taking into account.
My human, real-life decision theory can be summarised as "look at as many possible end-result worlds as I can, and at what actions will bring them into being; evaluate how much I like each of them; then figure out which actions are most efficient at leading to the best worlds". But that doesn't exactly fly when you're programming a computer, you need something that can be fully formalised, and that is where those strange Omega scenarios are useful, because your code must get it right "on autopilot", it cannot improvise a smarter approach on the spot - the formula is on paper, and if it can't solve a given problem, but another one can, it means that there is room for improvement.
In short, DT problems are just clever software debugging.
I agreed with everything you said after "I don't think so". So I am left confused as to why you don't think so.
You analogize DT problems as test data used to determine whether we should accept or reject a decision theory. I am claiming that our requirements (i.e. "reflective consistency") are so unrealistic that we will always be able to find test data forcing us to reject. Why do you not think so?
Decision theories should usually be seen as normative, not descriptive. How "realistic" something is, is not very important, especially for thought experiments. Decision theory cashes out where you find a situation that can indeed be analyzed with it, and where you'll secure a better outcome by following theory's advice. For example, noticing acausal control has advantages in many real-world situations (Parfit's Hitchhiker variants). Eliezer's TDT paper discusses this towards the end of Part I.
Because I suspect that there are only so many functionally different types of connections between events (at the very least, I see no hint that there must be infinitely many) and once you've found them all you will have the possibility of writing a DT that can't be led to corner itself into suboptimal outcomes due to blind spots.
Not to me. But a reference might repair that deficiency on my part.
See Eliezer's posts on Newcomb's Problem and regret of rationality and TDT problems he can't solve.
(Incidentally, I found those reference in about 30 seconds, starting from the LW Wiki page on Parfit's Hitchhiker.)
Ah! Thank you. I see now. The circumstance in which a CDT agent will self modify to use a different decision theory are that:
Well, ok, though I wouldn't have said that these are cases where CDT agents do something weird. These are cases where EYDT agents do something weird.
I apologize if it seems that the target of my sarcasm is you WrongBot. It is not.
EY has deluded himself into thinking that reflective consistency is some kind of gold standard of cognitive stability. And then he uses reflective consistency as a lever by which completely fictitious data can uproot the fundamental algorithms of rationality. Which would be fine, except that he has apparently convinced a lot of smart people here that he knows what he is talking about. Even though he has published nothing on the topic. Even though other smart people like Robin tell him that he is trying to solve an already solved problem.
I would say more but ...
This manuscript was cut off here, but interested readers are suggested to look at these sources for more discussion: Bibliography Gibbard, A., and Harper, W. L. (1978), "Counterfactuals and Two Kinds of Expected Utility", in C. A. Hooker, J. J. Leach, and E. F. McClennen (eds.), Foundations and Applications of Decision Theory, vol. 1, Reidel, Dordrecht, pp. 125-162.
Reflective consistency is not a "gold standard". It is a basic requirement. It should be easy to come up with terrible, perverse decision theories that are reflectively consistent (EY does so, sort of, in his TDT outline, though it's not exactly serious / thorough). The point is not that reflective consistency is a sign you're on the right track, but that a lack of it is a sign that something is really wrong, that your decision theory is perverse. If using your decision theory causes you to abandon that same decision theory, it can't have been a very good decision theory.
Consider it as being something like monotonicity in a voting system; it's a weak requirement for weeding out things that are clearly bad. (Well, perhaps not everyone would agree IRV is "clearly bad", but... it isn't even monotonic!) It just happens that in this case evidently nobody noticed before that this would be a good condition to satisfy and hence didn't try. :)
Am not sure that decision theory is an "already solved" problem. There's the issue of what happens when agents can self-modify - and so wirehead themselves. I am pretty sure that is an unresolved "grand challenge" problem.
TDT gets better outcomes than CDT when faced with Newcomb's Problem, Parfit's Hitchhiker, and the True Prisoner's Dilemma.
When does CDT outperform TDT? If the answer is "never", as it currently seems to be, why wouldn't a CDT agent self-modify to use TDT?
Because it can't find a write-up that explains how to use it?
Perhaps you can answer the questions that I asked here What play does TDT make in the game of Chicken? Can you point me to a description of TDT that would allow me to answer that question for myself?
Suppose I'm an agent implementing TDT. My decision in Chicken depends on how much I know about my opponent.
As Eliezer says here, the one-sentence version of TDT is "Choose as though controlling the logical output of the abstract computation you implement, including the output of all other instantiations and simulations of that computation."
I'm not sure this is right. Isn't there a correlated equilibrium that does better?
So TDT is different from CDT only in cases where the game is perfectly symmetric? If you are playing a game that is roughly the symmetric PD, except that one guy's payoffs are shifted by a tiny +epsilon, then they should both defect?
Thank you. I hope you realize that you have provided an example of a game in which CDT does better than TDT. For example, in the game with the payoff matrix shown below, there is a mixed strategy Nash equilibrium which is better than the symmetric cooperative result.
Do you have an example of a problem on which CDT or EDT does better than TDT?
I have yet to see a description of TDT which allows me to calculate what TDT does on an arbitrary problem. But I do know that I have seen long lists from Eliezer of problems that TDT does not solve that he thinks it ought to be improved so as to solve.
The world isn't sufficiently formalized for us to meet that standard for any decision theory (though we come closer with CDT and TDT than with EDT, in my opinion). However, cousin_it has a few recent posts on formalized situations where an agent of a more TDT (actually, UDT) type does strictly better than a CDT one in the same situation. I don't know of any formalization (or any fuzzy real-world situation) where the opposite is true.
I apparently misled you by using that word "arbitrary". I'm not asking for solutions to soft problems that are difficult to formalize. Simply solutions to the standard kinds of games already formalized in game theory. For example, the game of Chicken. Can anyone point me to a description that tells me what play TDT would make in this game? Or what mixed strategy it would use? Both assuming and not assuming the reading of each other's code.
ETA: Slightly more interesting than the payoff matrix shown in the wikipedia article is the case when the payoff for a win is 2 units, with a loss still costing only -1. This means that in the iterated version, the negotiated solution would be to alternate wins. But we are interested in the one-shot case.
Can TDT find a correlated equilibrium? If not, which Nash equilibrium does it pick? Or does it always chicken out? Where can I learn this information?
Since CDT and EDT don't solve those problems either, all this justifies saying is that TDT does better on some problems, and the same on others, not "worse on others".
For every possible decision theory, there is a "nemesis" environment - where it does extremely badly. That is no-free-lunch fall out.
A "nemesis" environment that feeds misleading evidence to a decision theory's underlying epistimology does not indicate the sort of problem illustrated by an environment in which a decision theory does something stupid with true information.
What you asked for was a case where a decision theory did worse than its rivals.
However, that seems pretty trivial if it behaves differently from them - you just consider an appropriate pathological environment set up to punish that decision theory.
Yes, in the context of Perplexed dismissing examples of TDT doing better than CDT because CDT was being stupid with true information.