Re:DeepMind says reinforcement learning is ‘enough’ to reach...
by tsimionescu on Thursday June 10, 2021 @08:15AM
(#27459756)
You are absolutely right about RL in practice. In fact, if we were to look at actual RL algorithms, I believe the paper's claims fall flat in many other ways. This is actually my criticism of it: its arguments are only convincing when RL is defined only as the extremely general notion of an agent seeking to maximize some reward function by interacting with an environment. This is so comically general that the only alternative I can think of is to posit a transcendental god.
Once we get into the details, their claims stop being iron clad. Even worse, some of their claims become actually hard or impossible to accept if applied to actual RL algorithms we have today. You give one good example with the difficulty of modeling the world. The implicit claim they make that this would be realizable in reasonable time (say, less than a billion years) is also not well supported. The idea that humans or mammals learn their social behaviors through RL rather than a good deal of reasoning from evolutionarily-trained first principles pretty clearly fails in the face of the poverty of the stimulus argument[0].
Overall, the claims in the paper tend to switch between obvious (if taken to talk about the general idea of maximizing reward) to almost certainly wrong (if taken to talk about known RL algorithms, reasonable time frames, and specific examples of what is supposed to be learned).
[0] the poverty of the stimulus argument may be controversial in linguistics where it was first formulated. Still, if applied to mammal or insect socialization, the extremely low time frames in which individuals of a species start exhibiting typical behaviors basically proves in my opinion that they are instincts, trained at the population level through evolution, not individual learning through RL. The extreme similarity of behavior between individuals of the same species, VS the variety of behaviors between different species, also suggests an important component of species-level rather than individual level learning.
Once we get into the details, their claims stop being iron clad. Even worse, some of their claims become actually hard or impossible to accept if applied to actual RL algorithms we have today. You give one good example with the difficulty of modeling the world. The implicit claim they make that this would be realizable in reasonable time (say, less than a billion years) is also not well supported. The idea that humans or mammals learn their social behaviors through RL rather than a good deal of reasoning from evolutionarily-trained first principles pretty clearly fails in the face of the poverty of the stimulus argument[0].
Overall, the claims in the paper tend to switch between obvious (if taken to talk about the general idea of maximizing reward) to almost certainly wrong (if taken to talk about known RL algorithms, reasonable time frames, and specific examples of what is supposed to be learned).
[0] the poverty of the stimulus argument may be controversial in linguistics where it was first formulated. Still, if applied to mammal or insect socialization, the extremely low time frames in which individuals of a species start exhibiting typical behaviors basically proves in my opinion that they are instincts, trained at the population level through evolution, not individual learning through RL. The extreme similarity of behavior between individuals of the same species, VS the variety of behaviors between different species, also suggests an important component of species-level rather than individual level learning.