Reward learning summary
A putative new idea for AI control; index here.
I’ve been posting a lot on value/reward learning recently, and, as usual, the process of posting (and some feedback) means that those posts are partially superseded already—and some of them are overly complex.
So here I’ll try and briefly summarise my current insights, with links to the other posts if appropriate (a link will cover all the points noted since the previous link):
I’m modelling humans as a pair , where is the reward function, and is a planning algorithm (called a planer) that maps rewards to policies. An agent is trying to learn .
The policy of a given human is designated . A pair is compatible if .
There is a no-free-lunch theorem for these pairs. Once the agent has learnt , it can get no further evidence from observing the human. At that point, any compatible pair is a valid candidate for explaining the human planner/reward.
Thus the agent cannot get any idea about the reward without making assumptions about the human planner (ie the human irrationality) - and can’t get any idea about the planner without making assumptions about the human reward.
Unlike most no-free-lunch theorems, a simplicity prior does not remove the result. LINK
It’s even worse than that: a simplicity prior can push us away from any “reasonable” . LINK
Ignoring “noise” doesn’t improve the situation: the real problem is bias. LINK
The formalism can also model situations like the agent “overriding” the human’s reward. LINK
There are Pascal’s mugging type risks in modelling reward override, where a very unlikely pair may still be chosen because the expected reward for that choice is huge. LINK
There is also the risk of an agent transforming human into rational maximisers of the reward its computed so far LINK.
Humans are not just creatures with policies, but we are creatures with opinions and narratives about our own values emotions, and rationality. Using these “normative assumptions”, we can start converging on better pairs. LINK LINK
Even given that, our values will remain underdefined, changeable, and open to manipulation. LINK
Resolving those problems with our values is a process much more akin to defining values that discovering them. LINK
There are important part of human values that are not easily captured in the formalism. LINK
- MIRI’s 2017 Fundraiser by (7 Dec 2017 21:47 UTC; 27 points)
- MIRI’s 2017 Fundraiser by (1 Dec 2017 13:45 UTC; 19 points)
- Towards an Axiological Approach to AI Alignment by (15 Nov 2017 2:07 UTC; 17 points)
- Formally Stating the AI Alignment Problem by (19 Feb 2018 19:06 UTC; 14 points)
- Evaluating Existing Approaches to AGI Alignment by (27 Mar 2018 19:57 UTC; 12 points)
- MIRI 2017 Fundraiser and Strategy Update by (EA Forum; 1 Dec 2017 20:06 UTC; 6 points)
- 's comment on Stable Pointers to Value II: Environmental Goals by (9 Feb 2018 9:59 UTC; 5 points)
I had idea for a prior for planners (the ‘p’ part of (p, R)) that I think would remove the no-free-lunch result. For a given planner, let its “score” be the average reward the agent gets for a randomly selected reward function (with a simplicity prior over reward functions). Let the prior probability for a particular planner be a function of this score, perhaps by applying a Boltzmann distribution over it. I would call this an evolutionary prior—planners that typically get higher reward given a randomly assigned reward function are more likely to exist. One could also randomize the transition function to see how planners do for arbitrary world-dynamics, but it doesn’t seem particularly problematic, and maybe even beneficial, if we place a higher prior probability on planners that are unusually well-adapted to generate good policies given the particular dynamics of the world we’re in.