A putative new idea for AI control; index here.

This post will present a simple model of an inconsistent human, and ponder how to resolve their inconsistency.

Let be our agent, in a turn-based world. Let and be two simple reward functions at each turn. The reward is thought of as being a 'long-term' reward, while is a short-term one.


Define as the agent's reward at turn (and similarly for ). Then, at turn , the agent has reward:

  • ,

with constants .

Essentially the and the have different discount rates, with the reward from fading much faster than than that of . Therefore the agent will be motivated to get the reward, but only if they can get this in the short-term. Sex, drugs, food, and many other pleasures often have these features (though they are, of course, much more complicated).

The inconsistency is that the human will continually reset their at each turn. If there were a single discount rate, that wouldn't a problem, as that would just scale the whole reward function, and reward functions, like utility functions, give the same decisions when scaled.

But with two discount rates, this is inconsistent. The agent will try and follow for long-term planning, but this will be disrupted if they encounter an along the way (and then presumably berate themselves for the lack of self-discipline). This can also be seen as a variant of the "humans are composed of multiple subagents" model, with corresponding to short-term greedy subagent.

So, we have a simple and not-completely-implausible model of an inconsistent human. The question is, how do we resolve it? None of the obvious approaches are ideal, but it's worth looking at their features.

Freeze the reward

This is the most obvious approach: simply freeze the reward, so that the reward at time is simply the same as reward at time (though, in the absence of time-travel, the rewards between and are no longer relevant).

This is the obvious approach; in practice, though, it will become equivalent with simply forgetting about entirely. After a few turns, the exponential shrinkage of the factor will make 's typical contribution insignificant. So this approach involves destroying one of 's sources of reward almost entirely.

Balance the rewards

Another approach would be to balance the rewards, set , either at the initial value of , the initial value of , or some other value.

This would make the the reward consistent, but has the opposite problem as the previous approach: the long-term importance of is now massively magnified relative to , so now the long-term plans will prioritise above much more than before.

Narrative/frequentist/feature based approach

Since we're not supposed to do this, let's anthropomorphise . We can imagine that has some sort of narrative about their existence -- they see themselves as being a certain type of person (possibly mainly connected with ), who has some quirks/indulgences/sins (possibly mainly connected with ).

If they want to extirpate entirely ("sin"), this is the same as the "Freeze the reward" approach. But they may instead prefer to live their life with roughly the same proportion of as before ("quirk"), or slightly less ("indulgence").

In that case, the new would be chosen for consequentialist reasons. Not by looking at the individual terms , , , and , but by looking at the consequences of following under "typical" circumstances, and designing the new reward to replicate this behaviour (and this distribution of reward features), while allowing more efficiency. This is, in itself, an interesting IRL problem. But it seems to make sense for humans, as we define ourselves a lot by what we do and experience, rather than by the pleasures and choices that lead up to those experiences.

New Comment
2 comments, sorted by Click to highlight new comments since:

Freezing the reward seems like the correct answer by definition, since if I am an agent following the utility function and I have to design a new agent now, then it is rational for me to design the new agent to follow the utility function I am following now (i.e. this action is usually rated as the best according to my current utility function).

This is, in itself, an interesting IRL problem. But it seems to make sense for humans, as we define ourselves a lot by what we do and experience, rather than by the pleasures and choices that lead up to those experiences.

Maybe? But this seems an after-the-fact ontological construction rather than necessarily tied to causative processes, so although it's a simple model that we should be able to align, I'm somewhat doubtful R_t describes human decision making with sufficient fidelity to be that useful to people wishing to change their axiology.