Maybe this is just being pedantic, and maybe you addressed it and I just missed it, but aren't search paradigms like novelty and interestingness still a form of objective search? The objective is just shifted to some other criteria. In the case of novelty to an unexplored location (or whatever criteria we're using for novelty) and interestingness definitely has some criteria, it just seems like we're relegating it to subconscious processes.
I don't think this is a pedantic response, and it is something I grappled with when writing this. One of the reasons for the impreciseness of my language in the post is because the book extrapolates beyond search as it might be formally defined in computer science, to include fields such as evolution, furniture design and the authors helping art students to find purpose. Once you're out of that space, holding on to the strict definitions is less valuable.
So if we move back into that space, one way of thinking about this is that novelty search is not improving an individual candidate but the archive as a whole. So novelty search appears objectiveless as it is ignoring the overarching objective. But it does have an objective - to map some candidate solutions to the problem, and to map them in such a way so as to maximise a sparseness criterion around each candidate solution (i.e., we want our candidate solutions to cover the space).
This is just basic novelty search, but we can give it far more complex forms. For instance, in Bayesian optimisation, we are looking through a search space to get some reward. If we use the upper confidence bound (UCB) as our acquisition function, we get a parameter β. We can set β to different values depending on how valuable we believe unexplored space to be. As we sample over time we might start with an initially high β value, which favours exploration, moving to a lower one, which favours exploitation in a high-reward area of the search space.
All of search, whether in computer science or outside of it, is fundamentally about dealing with a really large space and trying to find a way to narrow it down such that you find things you want to find. What the OMNI researchers are arguing is that interestingness might be a pretty good way to narrow it down, because it seems to concentrate search on things that humans believe is worthwhile and novel.
I guess your point is maybe something like - but we have an objective then: find worthwhile and novel things! But the difficulty is that “find worthwhile and novel things” is not something we necessarily know how to do. So OMNI attempts to approximate that judgement.
Let's say we have an algorithm that controls the guy in Minecraft. If we just ask the algorithm to do novel things, it might just do the following tasks:
- collect 1 log
- collect 2 logs
- collect 3 logs
- collect 4 logs
Interestingness is a way of saying, no, we want you to go and explore spaces and interact with creatures and complete the game, without having to program any of that manually. Instead, feeding millions of reddit posts and YouTube videos to a foundation model may have inadvertently programmed those ideas in as interesting (because people like to discuss things which are worthwhile and novel).
I was trying to combine that argument with the work in Stanley and Lehman's book to say that interestingness can be deployed in a number of places, particularly in the scientific process. Interestingness is not the only way to cut down the search space, and other ways may rely less on the subconscious.
For now, it seems as if there may be a trade-off between having a more formal search process and finding less useful things, and having a less formal search process and finding more useful things.
Many researchers, being expert at statistics, know that they just have to find the right metric and use the right hack for the desired answer to pop out!
Maybe this is just being pedantic, and maybe you addressed it and I just missed it, but aren't search paradigms like novelty and interestingness still a form of objective search? The objective is just shifted to some other criteria. In the case of novelty to an unexplored location (or whatever criteria we're using for novelty) and interestingness definitely has some criteria, it just seems like we're relegating it to subconscious processes.
I don't think this is a pedantic response, and it is something I grappled with when writing this. One of the reasons for the impreciseness of my language in the post is because the book extrapolates beyond search as it might be formally defined in computer science, to include fields such as evolution, furniture design and the authors helping art students to find purpose. Once you're out of that space, holding on to the strict definitions is less valuable.
So if we move back into that space, one way of thinking about this is that novelty search is not improving an individual candidate but the archive as a whole. So novelty search appears objectiveless as it is ignoring the overarching objective. But it does have an objective - to map some candidate solutions to the problem, and to map them in such a way so as to maximise a sparseness criterion around each candidate solution (i.e., we want our candidate solutions to cover the space).
This is just basic novelty search, but we can give it far more complex forms. For instance, in Bayesian optimisation, we are looking through a search space to get some reward. If we use the upper confidence bound (UCB) as our acquisition function, we get a parameter β. We can set β to different values depending on how valuable we believe unexplored space to be. As we sample over time we might start with an initially high β value, which favours exploration, moving to a lower one, which favours exploitation in a high-reward area of the search space.
All of search, whether in computer science or outside of it, is fundamentally about dealing with a really large space and trying to find a way to narrow it down such that you find things you want to find. What the OMNI researchers are arguing is that interestingness might be a pretty good way to narrow it down, because it seems to concentrate search on things that humans believe is worthwhile and novel.
I guess your point is maybe something like - but we have an objective then: find worthwhile and novel things! But the difficulty is that “find worthwhile and novel things” is not something we necessarily know how to do. So OMNI attempts to approximate that judgement.
Let's say we have an algorithm that controls the guy in Minecraft. If we just ask the algorithm to do novel things, it might just do the following tasks:
- collect 1 log
- collect 2 logs
- collect 3 logs
- collect 4 logs
Interestingness is a way of saying, no, we want you to go and explore spaces and interact with creatures and complete the game, without having to program any of that manually. Instead, feeding millions of reddit posts and YouTube videos to a foundation model may have inadvertently programmed those ideas in as interesting (because people like to discuss things which are worthwhile and novel).
I was trying to combine that argument with the work in Stanley and Lehman's book to say that interestingness can be deployed in a number of places, particularly in the scientific process. Interestingness is not the only way to cut down the search space, and other ways may rely less on the subconscious.
For now, it seems as if there may be a trade-off between having a more formal search process and finding less useful things, and having a less formal search process and finding more useful things.
If you'd like to read more:
- https://www.cs.swarthmore.edu/~meeden/DevelopmentalRobotics/lehman_ecj11.pdf
- https://direct.mit.edu/evco/article/32/3/249/116787/Preliminary-Analysis-of-Simple-Novelty-Search
Many researchers, being expert at statistics, know that they just have to find the right metric and use the right hack for the desired answer to pop out!
Quite sure I'm oversimplifying, but here's my takeaway. Take baby steps and reevaluate. Pay attention.
I would call this an Eureka moment.