4 ms·
"A sufficiently powerful and general reinforcement learning agent may ultimately give rise to intelligence and its associated abilities. ... We do not offer any
by Animats 5y ago
"A sufficiently powerful and general reinforcement learning agent may ultimately give rise to intelligence and its associated abilities. ... We do not offer any theoretical guarantee on the sample efficiency of reinforcement learning agents."
OK. This basically says "evolution works". But how fast? Biology took tens of millions of years to boot up.
An related question is how much compute power does evolution, viewed as a reinforcement learning system. have? That's probably something biologists have thought about. Anyone know? Evolution is not a very fast or efficient hill-climbing system, but there are a large number of parallel units. It's not a philosophical question; it's a measurable one. We can watch viruses evolve. We can watch bacteria evolve. Data can be obtained.
Two questions I pose occasionally are "how do we do common sense, defined as not screwing up in the next 30 seconds", and "why does robotic manipulation in unstructured situations still suck after 50 years". A good question to ask today is why reinforcement learning does so badly on those two problems. In both cases, you can define an objective function, but it may not be well suited to hill climbing.
- londons_explore 5y agoIn some cases biological 'genetic algorithm' hill climbing can be remarkably ineffective. For example, the classic "design a car that can drive over this terrain" problem, even after a billion generations (~ the same number as life on earth), shows no substantial performance improvement. That makes me suspect something is missing from our biological genetics model.
- wnkrshm 5y agoenergy supply and other constraints (material, robustness ...) are a good explanation though - an organism can't grow out of aluminum or steel
- EE84M3i 5y agoI suspect they're talking about this (sort of) simulation https://rednuht.org/genetic_cars_2/ https://rednuht.org/genetic_cars_2/
- kaba0 5y agoI think the number of parameters is remarkably (multiple orders of magnitude) different between even the simplest bacteria vs the model used in the car. And then genomics can also do some more advanced techniques like copy a whole gene and start modifying that, etc.
- 29athrowaway 5y agoWe just want AI to be able to think. We do not need an AI with an autonomic nervous system, or many of the functions in the central nervous system. We do not need AI to be very power efficient. If it takes several megawatts of electricity to get our first strong AI working, so be it. So, we do not have as many constrains as life did.
- ifdefdebug 5y agoDefine "think". And then prove it can be done without the kind of nervous system you say we don't need.
- 29athrowaway 5y agoThe embodied cognition folks believe that the embodiment of sentience affects cognition. That is probably true to some extent. I mean, if we make an AI that has an orgasm each time it blows up something with a hellfire missile, it will probably learn to find ways to blow up more things more frequently and efficiently. Our cognition is affected by pain, hunger, thirst, cold, heat, pleasure, smells, sounds, etc... positive and negative reinforcements.
- jacquesm 5y agoThat is the most impressive use of the word 'just' in a long time. Note that all the other bits are solved, and have been solved since the 60's. It's the 'just think' bit that has proven to be a little bit harder than we thought it would be.
- randcraw 5y ago> "why does robotic manipulation in unstructured situations still suck after 50 years" Great point. Until the promoters of RL can build us a robot that can 1) walk gracefully through a typical home that has stairs and closed doors, 2) cook a meal with pots and pans, and 3) get back up after it falls down -- I suggest we take their claims of impending Singularity with a big grain of salt.
- sbierwagen 5y agoConsider if Boston Dynamics released a video of a robot doing all three things tomorrow, and announced that they were taking orders for immediate shipment, priced at $100,000. How far away would you think the singularity would be then? Separately, a hostile or indifferent AI could still cause a heck of a lot of trouble for human civilization without the first two things. Consider an autofactory clearing room for expansion with bulldozers, no need to navigate stairs there. Bullets or smart glide bombs don't need to understand doorknobs. Etc.