DEEP REINFORCEMENT LEARNING
The mankind has always looked for an intelligent assistant.
Computers a byproduct of this exploration has reached the stage of deep reinforcement learning. It is radically different from supervised learning and takes computer / machine learning on a totally different trajectory. The aim to create artificial agents that can achieve human level of performance & generality. The agents, like humans, learn for themselves, to create successful strategies leading to best long term rewards.
This is the paradigm shift in learning. It happens through trial & error. It’s based solely on rewards & punishments and is known as reinforcement learning. Agents learn for raw inputs; vision is one example, domain heuristics & hand engineered features have no role to play. This is deep reinforcement learning. DeepMind an pioneering company of this research, is now an integral arm of Google.
The Artificial Intelligence landscape changed for good when this system of learning based AlphaGo, in October 2015, became the first program to defeat a professional Go player. Chinese game Go has infinite options. In March 2016, AlphaGo defeated Lee Sedol, strongest player of last decade with incredible 18 world titles, by four games to one. The match was watched by 200 million viewers. Deep reinforcement learning and consequent AI were there to stay.
Two years ago, introduced a widely successful algorithm; Deep Q-Networks, DQN. This algorithm stores all of the agent’s experiences. It then randomly samples & the experiences are replayed. The result is diverse & decorrelated training data. This is the core concept & the operation. DQN was extremely successful with Atari 2600 console. Deep reinforcement learning agents have been successful in variety of challenging games / tasks. The agents capabilities can be improved upon to make a positive impact on society, healthcare can be a notable area.
DEEP REINFORCEMENT LEARNING IS AT THE CORE OF WHAT AI PROMISES TO DELIVER