The Mostly "Living" Artificial Neural Networks
Kinda cool
Robots trying to walk and doing manual work have been the prime example of "tin machines" leaving humans jobless. One of the most prime examples of reinforcement learning is just the ones we were talking about. The title might be clickbait but it truly defines what I think of as one of the most beautiful approaches to Artificial Intelligence.
They are living in the sense that these neural networks show a dynamic behavior. RL short for reinforcement learning tries to maximize the utility function or reward in general terms. RL agents work for the sole purpose of living and increasing their treats most like we humans and other beings do. They change what they learn and how they learn and iterate versions of themselves to survive in the environment we put them in. RL showed promise even before the deep learning models got famous. Even when multi-layered neural networks were not in much use these helped to do tasks that would have required large amounts of computing through other artificial intelligence approaches. They are used in literally any way one can imagine(mostly) , from Robots to problem-solving, Self-driving cars to Natural language processing.
One of the biggest feats of RL is that it was able to defeat the Go world Champion in 2016 about which even its creators were not sure. AlphaGo became the world Go champion. Go is one of the most complex games and is harder in orders of magnitude than chess. Go has 2.1*10^170 possible positions which are even larger than the no. of atoms of our universe. Even with all the world’s computers combined, we can not calculate every position in a human lifetime. Earlier than AlphaGo, the heuristics approach was used for building AI systems that could play Go but they were not that successful as most would plateau at some point and that point wasn't even near a world champion but one of the most fascinating things about AlphaGo and other RL agents is there is just no bar of how much they can master and learn. This is one of the most inspiring achievements of AI. Some say that Lee Sedol might be the only one ever who has defeated AlphaGo in even one match.
One of the most common starting points in RL is balancing a cart pole in simulation. I will leave the link here for you to try a google collab
Will leave resources in the last if dear reader you want to learn more.
One of the most fascinating use cases of RL is Robot control. RL is kinda the mind behind a robot. My favorite AI Lab OpenAI.com taught a robot arm to solve a Rubik's cube
When you finish watching the video you probably are in awe but the possibilities of RL are endless.
Reinforcement Learning is one of the sub-topics of Machine Learning where an agent is situated in an environment where a reward is given for certain learning behaviors. It combines some great ideas of both supervised and unsupervised learning. One doesn't need to label the data or supervise the agent. Some great outcomes have been observed when there was no supervision and the agent was let to roam free and interact with the environment. Some very complex and high accuracy computer games like Dota 2 have been mastered and world champions defeated by RL agents. This approach is not limited to games. It plays a big role in the Self-driving sector and even has better predict outcomes for molecule simulations.
One can argue that RL is the AGI but when these agents interact they nowhere show the level of intelligence of an organism let alone Humans but the hope is someday they will and it would have some great outcomes.
PS: I am truly grateful to you for reading what I write. These are just my thoughts and nothing else. I read and gather information about the stuff that I like and seems fascinating to me and then I write just for fun but thank you for making it worthwhile.
If you want to see a blog post on the topic you like, buy me a book . And will just leave my eth address for fun cause it seems cool. eth - 0x5c0265DCFBeeA7A363098483fE4910052D4DD668
Thank You.
And try Following Pieter Abbeel he is one the most well-known researchers in the field and David Silver too.