If an agent can sense, think, and act in an unknown environment, how MUST it act? Reinforcement Learning (RL) is a general paradigm under which an agent, by trial and error, can discover actions that result in long-term gain. Integrating ideas from various disciplines, RL has matured as a field of study in the last few decades. Its successes range from training programs to decrease the waiting time for elevators in a building; to maximising profits from stock-trading; and to numerous tasks in robotic control and decision-making.
Historically, the gap between the promise of RL and its practical effectiveness has been due to the lack of suitable representations (and representation-discovery mechanisms) in domains of interest. Interestingly, it is precisely this gap that the emerging paradigm of deep learning is beginning to fill. Deep learning is a data-driven technique that is capable of learning complex non-linear input-output relationships, which are represented using neural networks with a large number of layers (hence "deep"). The combination of RL with deep learning has registered remarkable successes in recent years. Notably, it has resulted in AI-based agents that exceed human-level performance on a suite of ATARI console games, and also on the more challenging game of Go.
I will begin this talk with an overview of RL and its key algorithmic elements. Specifically, I will outline the role of representations in determining the success of RL. I will follow with a brief survey of deep learning, and proceed to describe the architecture of the agents that have recently succeeded at ATARI and Go. I will end the talk with thoughts about this new era for "deep reinforcement learning".