Quick notes on Reward is enough. I wasn’t aware this paper was supposed to be controversial, but it sparked quite a lively debate on Learning Salon. To me the the paper makes perfect sense. Argues that general intelligence arises in complex environments, even from seemingly innocuous signals. Wasn’t some version of this written by Skinner in the mid-1900s? Aren’t economists usually writing these types of papers, substituting intelligence for some other behavior that’s supposed to emerge from a model?
Anyhow. My TL;DR: intelligence is emergent behavior facilitated by any kind of reward present in a complex environment. Titled “Reward is Enough” for likely PR reasons, actually argues to say reward is a necessary but insufficient condition.
The behaviour of the squirrel may be understood as maximising a cumulative reward such as satiation (i.e. negative hunger). In order for a squirrel to minimise hunger, the squirrel-brain must presumably have abilities of perception (to identify good nuts), knowledge (to understand nuts), motor control (to collect nuts), planning (to choose where to cache nuts), memory (to recall locations of cached nuts) and social intelligence (to bluff about locations of cached nuts, to ensure they are not stolen).
Well yeah, these behaviors seem distinct because they named them as such. If I granulate on a “nut-usefulness” level then perception, knowledge, planning, memory and social intelligence can all be a single ability.
Knowledge: information that is internal to the agent. May be learned or innate; long-lived environments favor the former.
innate understanding of predator evasion may be necessary before there is any opportunity to learn this knowledge. Note, however, that the extent of prior knowledge is limited both in theory (by the capacity of the agent) and in practice (by the difficulty of constructing useful prior knowledge).
Perception could be subserving e.g. maximization of healthy food, avoiding accidents, minimizing pain. Social intelligence could arise just from the presence of other agents. They kind of relate language to social intelligence, saying it arises from an agent’s need to communicate its complex environment to achieve goals.
I guess a simplified view of the perception hierarchy goes something like:
- Survival
- Energy
- Food acquisition
- Discern edible items
- Perception
- Discern edible items
- Minimize pain
- Avoid hurtful situations (e.g. falling off a cliff)
- Perception
- Avoid hurtful situations (e.g. falling off a cliff)
- Food acquisition
- Energy
Then some not-too-convincing lamentation about generalization being useful for maximizing an agent’s understanding within a rich environment. I would’ve claimed that an agent perhaps has limited cognitive capacity, and as such generalization is a consequence of interference but what do I know. Maybe I don’t get what they’re trying to say.
We do not offer any theoretical guarantee on the sample efficiency of reinforcement learning agents. Indeed, the rate at and degree to which abilities emerge will depend upon the specific environment, learning algorithm, and inductive biases.
Which environment? Depends on what you’re trying to achieve. But it doesn’t matter, because the real question is… Which reward signal? The desire to manipulate the reward signal often arises from the idea that only a carefully constructed reward could induce general intelligence. This paper suggests that, the more complex the environment, the more complex the behavior, irrespective of reward (e.g. pebble collecting).
In their framework, reproductive success is e.g. a possible reward that drives the emergence of intelligence. Reinforcement learning is the best choice, because:
- unsupervised learning and prediction lack action selection
- supervised learning is a shortcut that can teach a subset of what humans know
- maximization of free energy or minimization of surprise can’t be directed to achieve specific goals across different environments
As such, they argue against offline RL (no feedback to direct the agent towards their particular goals).
Is the reward signal too impoverished? I think they agree with me on this one, the single learning signal will actually lead to the emergence of “sub-signals” (e.g. survival necessitates food acquisition) which the agent will learn:
Many solution methods, including model-free reinforcement learning, learn to associate future reward with features of observations, through value function approximation, which in turn provide rich secondary signals that drive learning of deeper associations through a recursive bootstrapping process.
Lest I forget: there is no distinction being made between “useful” and “intelligent”.
Asking a self-driving car to bring you closer to a specific location, even in the presence of a perfectly complex environment, may lead to both intelligent and “wrong” behavior, failing to deliver on the promise of AI.
But I guess the same is true with humans, so eh. I’m sure Yudkowski would have some comments there.
Some kind of summary: Under the “reward is enough” framework, every intelligent agent is wired to follow a simple reward but the behavior that emerges is largely influenced by priors (“model weights”) and environment interactions (“gradient descent”). Thus it’s just shifting the burdern of explaining intelligence towards other parts of the system.