|
One caveat: It is unlikely that you can min/max more than one variable in a system at once with a single pertubation. As an example: Say you are trying to give out bank loans to white and black people, and to men and women. You have a metric that you give loans out: namley their credit score. You are not guaranteed to be able to give out bank loans to all people fairly. Say that black women, via the randomness of life, have lower credit scores than white women. So, because you give out loans via credit scores, you disadvantage black women. This is not okay. So, you then include some factor in the bank loaning process that accounts for that the race of the person applying for the loan. This then corrects the racial disparity in the loans to balck and whote women. However, what did you do to the rates of applications to black men and white men? Your new metrics have likely affected those rates, maybe negatively affecting women as a whole as comapred to men. Maybe it results in lowered rates to black men as comapred to white men. Who knows. You repeat this process of sdjustment until all loans are given out fairly. Then you look at what was done to Native American populations, homosexuals, asians, poor people, etc. I'm not saying that you are forbidden from being able to iteratively get to fairness in bank loans, but a single rule change is very unlikely to result in total fairness amoung all possible divisions of people. Additionally, to get to fairness in all bank loans, the rule set is likely to be somewhat complex [0] and then it'll have to change with time. [0] The handshake problem maybe? n(n-1)/2 complexity? With just 7 racial categories alone you get 21 total interactions that need to be adjusted for, Add in 3 gender categories and you then have 210 interactions. Now put in 4 monetary classes, you get 3486 interactions to deal with. Now add in pregnant people, veterans, the disabled, political views, religious groups, etc. It roughly grows as N^2 and gets hairy really quick. |