4 ms·

I'm curious, care to elaborate?

by mrwh 4y ago

wyager 4y ago
Rawls' proposed social optimization strategy is basically, implicitly, minimaxing (although I doubt he would describe it in similar terms), which is terrible for social outcomes. The return/variance pareto frontier for societal optimization is super sharp, so minimaxing makes almost everyone way worse off than expected utility maximization. Even if you're critical of EV maximization in general, it's really hard to claim that minimaxing performs acceptably at society-scale.
- whimsicalism 4y ago
  Uh, source for your claim? I think that diminishing marginal utility of wealth probably means that EV maximization and so-called "maximin" overlap a considerable amount of the time.
  - wyager 4y ago
    The source is common sense, experience, and original thought. I don't copy every thought I have from an NYT article or whatever you're looking for. Minimax is emphatically not even remotely close to EV maximization, especially for high-dimensional optimization spaces with sharp risk/return curves as I described.
    - whimsicalism 4y ago
      You're making empirical claims about the world as if they are obvious, so I expected you would have a shred of evidence backing what you were saying, but I guess I forgot we were on HN. Unsure why you would think I was expecting a NYT article (except as some weird cultural signaling point?), I was thinking more along the lines of a welfare economics or econometrics journal article.
      - wyager 4y ago
        The two likely possibilities are that you have enough background in utility/optimization/portfolio/whatever theory to reverse-engineer my thought process, or you need a lot of additional background that I can't feasibly provide in a single HN comment. Either way, linking to some random economics journal article is unlikely to help. I might as well link to a bunch of textbook pages defining all the words I used. As I said, this is an original thought of mine, so if an article reiterating what I said exists, I am in not much better a position to find it than anyone else on here. Apologies if this seems harsh, but people's habit of asking for a source on every original or propositional claim on this website really bugs me. Sometimes people are sharing new observations!
        whimsicalism 4y ago
        I have a background in economics and optimization and have not seen a claim as strong as the one you are making. Again, given that you are making an empirical claim about the world as we find it, I would expect you would be basing that on something you could point to - but I guess not. Not going to keep responding, I've said all I wanted to.
        tomcatfish 4y ago
        Look above you in the thread[1] for a very simple demonstration of how "EV maximization and so-called "maximin" overlap a considerable amount of the time" can be false. I am not really clear on what background you could have such a background but be unable to generate this fact, Wikipedia seems to suggest it strongly when it points out that "In non-zero-sum games, this is not generally the same as minimizing the opponent's maximum gain, nor the same as the Nash equilibrium strategy" [2]. This isn't a personal attack, but you really could have got this one solved by looking at either the thread or Wikipedia, but chose to assign someone else research homework instead. If you're curious, check! It's really fun and you learn a lot. [1]: https://news.ycombinator.com/item?id=31588384 https://news.ycombinator.com/item?id=31588384 [2]: https://en.wikipedia.org/wiki/Minimax#Maximin https://en.wikipedia.org/wiki/Minimax#Maximin
- mrwh 4y ago
  So for a specific example, in a world in which (say) 50% of people earn a dollar a day, and 50% a hundred dollars a day, and we are judging two potential policies: 1), that reduces the number of people on a dollar a day to 25%, so that now 75% of people earn a hundred dollars a day; or 2) that means that no one is an a dollar a day, but that everyone is now on 50 dollars a day. Rawls would say go to option 2, where plenty of people are worse off, but now no one is terribly off. Is that a fair example? Net, the situation is worse; but now no one is destitute.
  - guerrilla 4y ago
    > Net, the situation is worse; but now no one is destitute. How do you measure this? I think if you could measure subjective suffering, the situation may be better. Did you assume that people needed the extra $50 for something?
  - wyager 4y ago
    Rawls would definitely pick the second option, but utilitarianism may or may not favor the first option (depends e.g. on dollar-utility mapping).
  - mrwh 4y ago
    To be clear, personally I think a world in which everyone has less but no one has none is, on a moral level, a better world. If this is a fair example of Rawls' thought, then I'm with Rawls, and am not a utilitarian.
  - throwaway8582 4y ago
    Sure, the values you picked for your example make the Minimax strategy look reasonable. You can just as easily choose values which make it look silly: 1) 1% of people earn $1 per day, and 99% of people earn $100 per day 2) Everyone earns $1.01 per day I think most people, myself included, would take their chances with the first option. The Veil of Ignorance is a useful thought experiment, but I don't think many people would consider a simple Minimax algorithm to be reasonable, and would take into account the well-being of the average person as well.
- slaymaker1907 4y ago
  I agree, but only to the extent that we need to consider a cutoff somewhere for extraordinary bad luck. Maybe we only consider maximizing the well-being of the 99.9% or the 99.99% of people like how we mostly ignore the maximum response time for a server in favor of looking at 95th percentile. Do you have any evidence for the claim that minmaxing causes a steep drop off in overall expected utility? I'm not sure if research has been done about this with traffic lights, but I think that could be a good case study about this. Minmax for traffic light timings would mean minimizing the maximum wait time at a traffic light, though I could see how such a strategy might hurt the overall throughput of the intersection. In the extreme case, you have one granny trying to cross a 6-lane stroad during rush hour; in the pure throughput model, the road would never stop for granny and in the minmax model, she would only need to wait about as long as it takes her to cross the road. Obviously, different models would still allow granny to cross the road eventually and maybe with less of a throughput hit than pure minmax. All of this is speculation though and I'm curious what the throughput hit would be for minmax.
  - wyager 4y ago
    > Do you have any evidence for the claim that minmaxing causes a steep drop off in overall expected utility? My evidence is just experience with optimization/utility theory. If you're looking for a paper describing this one takeaway out of thousands from this body of theory, I'm not sure what it's called. In general, any strategic constraint reduces EV, unless the constraint happens to already be a component of the optimal strategy. That doesn't tell you how steep the dropoff is, just that it exists. Determining the variance/EV tradeoff precisely depends entirely on the structure of the optimization space. However, any complex optimization space exhibits essentially the same behavior; crushing the outcome distribution to be super narrow (latency, wealth, ROI, test scores, etc.) is typically only possible at great cost to the median and average cases.

I'm curious, care to elaborate?

by mrwh 4y ago

I'm curious, care to elaborate?

Rawls' proposed social optimization strategy is basically, implicitly, minimaxing (although I doubt he would describe it in similar terms), which is terrible for social outcomes. The return/variance pareto frontier for societal optimization is super sharp, so minimaxing makes almost everyone way worse off than expected utility maximization. Even if you're critical of EV maximization in general, it's really hard to claim that minimaxing performs acceptably at society-scale.

whimsicalism 4y ago
Uh, source for your claim? I think that diminishing marginal utility of wealth probably means that EV maximization and so-called "maximin" overlap a considerable amount of the time.
- wyager 4y ago
  The source is common sense, experience, and original thought. I don't copy every thought I have from an NYT article or whatever you're looking for. Minimax is emphatically not even remotely close to EV maximization, especially for high-dimensional optimization spaces with sharp risk/return curves as I described.
  - whimsicalism 4y ago
    You're making empirical claims about the world as if they are obvious, so I expected you would have a shred of evidence backing what you were saying, but I guess I forgot we were on HN. Unsure why you would think I was expecting a NYT article (except as some weird cultural signaling point?), I was thinking more along the lines of a welfare economics or econometrics journal article.
    - wyager 4y ago
      The two likely possibilities are that you have enough background in utility/optimization/portfolio/whatever theory to reverse-engineer my thought process, or you need a lot of additional background that I can't feasibly provide in a single HN comment. Either way, linking to some random economics journal article is unlikely to help. I might as well link to a bunch of textbook pages defining all the words I used. As I said, this is an original thought of mine, so if an article reiterating what I said exists, I am in not much better a position to find it than anyone else on here. Apologies if this seems harsh, but people's habit of asking for a source on every original or propositional claim on this website really bugs me. Sometimes people are sharing new observations!
      - whimsicalism 4y ago
        I have a background in economics and optimization and have not seen a claim as strong as the one you are making. Again, given that you are making an empirical claim about the world as we find it, I would expect you would be basing that on something you could point to - but I guess not. Not going to keep responding, I've said all I wanted to.
        tomcatfish 4y ago
        Look above you in the thread[1] for a very simple demonstration of how "EV maximization and so-called "maximin" overlap a considerable amount of the time" can be false. I am not really clear on what background you could have such a background but be unable to generate this fact, Wikipedia seems to suggest it strongly when it points out that "In non-zero-sum games, this is not generally the same as minimizing the opponent's maximum gain, nor the same as the Nash equilibrium strategy" [2]. This isn't a personal attack, but you really could have got this one solved by looking at either the thread or Wikipedia, but chose to assign someone else research homework instead. If you're curious, check! It's really fun and you learn a lot. [1]: https://news.ycombinator.com/item?id=31588384 https://news.ycombinator.com/item?id=31588384 [2]: https://en.wikipedia.org/wiki/Minimax#Maximin https://en.wikipedia.org/wiki/Minimax#Maximin
mrwh 4y ago
So for a specific example, in a world in which (say) 50% of people earn a dollar a day, and 50% a hundred dollars a day, and we are judging two potential policies: 1), that reduces the number of people on a dollar a day to 25%, so that now 75% of people earn a hundred dollars a day; or 2) that means that no one is an a dollar a day, but that everyone is now on 50 dollars a day. Rawls would say go to option 2, where plenty of people are worse off, but now no one is terribly off. Is that a fair example? Net, the situation is worse; but now no one is destitute.
- guerrilla 4y ago
  > Net, the situation is worse; but now no one is destitute. How do you measure this? I think if you could measure subjective suffering, the situation may be better. Did you assume that people needed the extra $50 for something?
- wyager 4y ago
  Rawls would definitely pick the second option, but utilitarianism may or may not favor the first option (depends e.g. on dollar-utility mapping).
- mrwh 4y ago
  To be clear, personally I think a world in which everyone has less but no one has none is, on a moral level, a better world. If this is a fair example of Rawls' thought, then I'm with Rawls, and am not a utilitarian.
- throwaway8582 4y ago
  Sure, the values you picked for your example make the Minimax strategy look reasonable. You can just as easily choose values which make it look silly: 1) 1% of people earn $1 per day, and 99% of people earn $100 per day 2) Everyone earns $1.01 per day I think most people, myself included, would take their chances with the first option. The Veil of Ignorance is a useful thought experiment, but I don't think many people would consider a simple Minimax algorithm to be reasonable, and would take into account the well-being of the average person as well.
slaymaker1907 4y ago
I agree, but only to the extent that we need to consider a cutoff somewhere for extraordinary bad luck. Maybe we only consider maximizing the well-being of the 99.9% or the 99.99% of people like how we mostly ignore the maximum response time for a server in favor of looking at 95th percentile. Do you have any evidence for the claim that minmaxing causes a steep drop off in overall expected utility? I'm not sure if research has been done about this with traffic lights, but I think that could be a good case study about this. Minmax for traffic light timings would mean minimizing the maximum wait time at a traffic light, though I could see how such a strategy might hurt the overall throughput of the intersection. In the extreme case, you have one granny trying to cross a 6-lane stroad during rush hour; in the pure throughput model, the road would never stop for granny and in the minmax model, she would only need to wait about as long as it takes her to cross the road. Obviously, different models would still allow granny to cross the road eventually and maybe with less of a throughput hit than pure minmax. All of this is speculation though and I'm curious what the throughput hit would be for minmax.
- wyager 4y ago
  > Do you have any evidence for the claim that minmaxing causes a steep drop off in overall expected utility? My evidence is just experience with optimization/utility theory. If you're looking for a paper describing this one takeaway out of thousands from this body of theory, I'm not sure what it's called. In general, any strategic constraint reduces EV, unless the constraint happens to already be a component of the optimal strategy. That doesn't tell you how steep the dropoff is, just that it exists. Determining the variance/EV tradeoff precisely depends entirely on the structure of the optimization space. However, any complex optimization space exhibits essentially the same behavior; crushing the outcome distribution to be super narrow (latency, wealth, ROI, test scores, etc.) is typically only possible at great cost to the median and average cases.