It is not surprising that many of the methods and algorithms explored in these lessons have some similarities to how humans perceive intelligence. After all, if it is humans that are building these methods, it is expected that we will create this with our bias of how we see the world. One of the most interesting similarities is how reinforcement learning algorithms can maximize the expected future reward over a long term horizon. This is perhaps what makes humans the most intelligent creatures on earth, we plan, execute and even sacrifice high short-term rewards for even smaller rewards given multiple times in the future. This is truly amazing.
Another obvious and important similarity is that of learning by trial and error. It is true that humans learn with supervision as well. But lots of the learning come from trial and error. There is no way of teaching a toddler to walk with pen and paper, or by reading books, a child will learn to walk by walking. As incredible as it sounds, it is a fact.
However, not everything is immediately obvious, and it is in fact, a mistake to just try to recreate a human brain. Usually, technology advances more quickly when we build systems that enhance humans, instead of trying to replace us. One of the fundamental differences between the way reinforcement learning works and how humans behave seem to be a lot the reward system. The fact that different humans can perceive the same reward signal different. In reinforcement learning, the reward signal is given by the environment, but it is not clear that this is how to world is actually model in reality. Sure, no human would consider that stepping on a nail is a positive signal, but it is true that many humans, especially, successful ones, usually have a way of "bending" reality to always look for the positive. Perhaps, this is something researchers should be working on these days.
- Reinforcement Learning, High-Level Cognition, and the Human Brain
- A Comparison of Human and Agent Reinforcement Learning in Partially Observable Domains
- Intrinsically Motivated Reinforcement Learning
- Intrinsic Motivation For Reinforcement Learning Systems
- Hierarchical Deep Reinforcement Learning: Integrating Temporal Abstraction and Intrinsic Motivation