Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[liopoil]

What are TDT and CDT?

https://en.m.wikipedia.org/wiki/TDT and https://en.m.wikipedia.org/wiki/CDT are not helping me here.

https://wiki.lesswrong.com/wiki/Causal_Decision_Theory
https://wiki.lesswrong.com/wiki/Timeless_decision_theory

@liopoil Have you read the linked paper on TDT?

I'd read that article a year or so ago and didn't really like it then, but I reread it just now and still didn't. Are you talking about the 120-page TDT paper by Yudkowsky? I'm not about to read the whole thing because I'm not sure I'm exactly on board with the premise. If there's a particular section you think I should look at I will.

My issue starts with Newcomb's "Paradox". I'm a two-boxer. The article states that "Omega is an almost perfect predictor of human behavior". OK, so I'm taking this to mean that I still am ultimately the one who makes the choice, and there's a non-zero chance (one in a billion maybe) that Omega is wrong. In this case, Omega surely knows that I am a two-boxer and so I am only going to get $1000, while sure, some TDT agent gets a million. But the TDT agent could have gotten a million AND a thousand! They made the wrong choice, while I made the best choice I could given the circumstance (yielding an expected value of $1000 + 1/1000, if there's a one-in-a-billion chance that Omega is wrong). I'm at a disadvantage because I'm a two-boxer, and yes, I would do better as a one-boxer (so if I could truly choose to be a one-boxer (TDT) before the game started, I would). But I do not have this option - I know that once the game has started I'll be a CDT agent again (one-boxer). I mean, it's called "timeless", but the game has a very clear timeline.

Ok, you are right that there is a difference - a game full of TDT agents wouldn't result in truthful alignment claims. But it does still cause problems. Say that it did get to a scenario where there's a claimed TDT survivor who has the option to defect to mafia and clinch a win. The mafia beg the TDT agent to take the win with them, but instead the survivor stays with town and lynches mafia. That night, some freak night actions occur and the survivor dies, and town eventually wins. The TDT agent clearly played against their own wincon, and mafia had a clear win stolen from them. That ruins games. This could also happen with prisoner-dilemma-like scenarios at the end of multiball games, or in a game of diplomacy. Especially in diplomacy, so many games are thrown by players who didn't played like TDT agents.

The premise of all games (mafia, Newcomb's, prisoner's dilemma, multiplayer dominion, diplomacy,...) is that all players play to maximize their utility at all times. TDT agents may outperform CDT overall in many cases, but adopting TDT is not really an option. If the survivor could sign a contract saying "I swear I won't defect to mafia at lylo", and have that contract be binding, that would be one thing. The CDT agent would also sign that contract. But such contracts are deliberately NOT part of the game. The survivor's win chances are hurt by their ability to defect, and there's nothing they can do to remove that ability.

I remember hearing the best example of this like 10 years ago, when I was 8 years old. If you're playing a game of chicken (two cars driving towards each other, each trying to make the other swerve first before crashing), then the best strategy is to tear your steering wheel off of the car and throw it out the window. This shows the other car that you really are an idiot and absolutely are not going to swerve - now you can't! The other car has no choice now but to swerve, and you win. I thought that was pretty great. But unless it's specifically a mechanic in a game, you aren't allowed to tear the steering wheel off your car, and the other players can assume that there is a point when you will swerve. Ok that analogy wasn't as good as I thought it would be but I still like it.

At this point I'm realizing that I'm more or less claiming that games, almost by definition, are played by CDT agents. I'm fine with that.

[AndrewisFTTW]

Oh God. The nerdiness. It hurts me.

[Galzria]

Oh God. The nerdiness. It hurts me.

It was glorious.

[ashersky]

Oh God. The nerdiness. It hurts me.

The best part of the whole thing was that he explained what playing chicken is.

[Calamitas]

Liopoil, you shouls really read the 120 pages (or at least the beginning). It's basically written for CDT agents and I believe that it can actually help you at achieving better results.

[liopoil]

Oh God. The nerdiness. It hurts me.

Uh, ok.

Oh God. The nerdiness. It hurts me.

The best part of the whole thing was that he explained what playing chicken is.

See, at first I was sure everyone knew what playing chicken was. But then I asked myself whether I knew what it was, and I wasn't totally sure I remembered right, so I had to look it up to check. At that point I figured I better be safe and include it, oh well.

Liopoil, you shouls really read the 120 pages (or at least the beginning). It's basically written for CDT agents and I believe that it can actually help you at achieving better results.

Ok, I'm ~35 pages in, reading quickly since it really doesn't seem like the content so far needs to fill 35 pages. Might read some more later.

I am perfectly fine with everything that I've read so far. In fact, it's made several of the same points I've made so far. I particularly liked the line:

Isaac Asimov once said of Newcomb’s Problem that he would choose only box A

But nothing so far changes anything I've said already. I'd still two-box, and I still stand by:

In general I don't think there is many scenarios where survivor should claim survivor, unless to immediately win with mafia. Survivors should just play like town really.

If someone is established as a TDT agent it makes very very much sense to claim; otherwise not.

TDT agents make games less fun in my book, but OK.

From the long abstract: "an evidential or causal decision-maker capable of self-modifying actions, given a choice between remaining an evidential or causal decision-maker and modifying itself to imitate a timeless decision-maker, will choose to imitate a timeless decision-maker on a large class of problems."

I don't dispute this at all. But the key part, "capable of self-modifying actions", doesn't apply here. I can't make myself a TDT agent before playing Newcomb's paradox, just like I can't sign a contract promising not to win with mafia as survivor or throw my steering wheel out the window. And in general I think games are designed so that players must make decisions causally, though as the paper says they will imitate a timeless decision-maker if given a mechanic to do so.

So I'm not sure that we have much of a disagreement here, besides that I don't think it's credible to claim to be an actual TDT agent in a mafia game. I remember you tried to do something like this in the "math games" thread while that was still a thing. Those games didn't really have a mechanic for self-modification either.

[Calamitas]

I'm a TDT agent everywhere, I have modified to one.