Intelligent Agent Foundations Forumsign up / log in
Stratified learning and action
post by Stuart Armstrong 29 days ago | discuss

A putative new idea for AI control; index here.

It seems that stratification is more flexible than I initially thought.

That’s because the default action or policy \(\emptyset\), which I was envisioning as a null action (or maybe the AI turning itself off) can actually be more than that. For instance, \(\emptyset\) could be an obsessive learning policy, for learning human values, for instance - and these human values can form the core of the AI’s value function \(\mathbf{W}\).

Then, stratification means that the AI will act to maximise human values, while estimating those values in accordance with what it would have calculated, had it been a pure value-estimator. This avoids the tension between value-learning and value maximising that bedevils most value-learners.



NEW LINKS

NEW POSTS

NEW DISCUSSION POSTS

RECENT COMMENTS

Can you provide links to the
by Vadim Kosoy on Two Questions about Solomonoff Induction | 0 likes

And I just wanted to write a
by Vadim Kosoy on Online Learning 1: Bias-detecting online learners | 0 likes

Also see the notion of
by Paul Christiano on Online Learning 1: Bias-detecting online learners | 2 likes

Given that this is my first
by Ryan Carey on Online Learning 1: Bias-detecting online learners | 1 like

I initially played around
by Devi Borg on Logical Inductors that trust their limits | 2 likes

I still feel like I don't
by Devi Borg on Logical Inductors that trust their limits | 2 likes

Running the traders on some r
by Sune Kristian Jakobsen on Variations of the Garrabrant-inductor | 0 likes

1. Note that IRL is
by Jessica Taylor on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

Stuart did make it easier for
by Patrick LaVictoire on (C)IRL is not solely a learning process | 0 likes

Nicely done. I should have
by Sam Eisenstat on The set of Logical Inductors is not Convex | 1 like

Ok, I think we need to
by Stuart Armstrong on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

I strongly predict that if
by Jessica Taylor on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

Wireheading the human is the
by Stuart Armstrong on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

Re 1: There are cases where
by Jessica Taylor on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

1. I don't really see the
by Stuart Armstrong on Heroin model: AI "manipulates" "unmanipulatable" r... | 0 likes

RSS

Privacy & Terms (NEW 04/01/15)