In this post we will discuss the importance of Reinforcement Learning for building an artificial intelligence.
Deep Learning is fundamental to function approximation nowadays. It can also be viewed as high-dimensional interpolation between data points.
We can do a lot of cool things like recognize or create samples from a distribution.
While this is an amazing tool in a self learning system it does not provide any real framework for an Artificial General Intelligence (AGI).
AGI describes a true artificial intelligence. It can learn any task a human or animal can learn. This is obviously beyond the scope of simple interpolation.
Markov Decision Process
This framework is provided by Reinforcement Learning. Especially by the Markov Decision Process (MDP).
MDP contains a state space , an action space , a probability function of transition from state to state after executing action and a reward function .
Now that we have defined our model we can think about it and see where it can be applied.
Of course a game of Tic-Tac-Toe can be seen as a MDP, but if you really apply the concept, one can describe reality as a MDP.
The state space is the universe and the action space contains any action a human can do. While we can’t see the true state of our surroundings, we humans create an *observation* from our senses.
Since it works for us and humans are the most general intelligent agent we know, MDP is an excellent framework to train an AGI.
Reinforcement Learning + Deep Learning = Artificial General IntelligenceDavid Silver
This framework combined with the ability of deep learning to approximate functions well given enough data makes for a powerful combination. This allows us to learn state-action values or directly a policy.
Since the goal in Artificial General Intelligence is to develop an intelligent agent, Markov Decision Processes provides the clear framework for it.
Combined with Deep Learning we really amp up the power of our agents.
Deep Reinforcement Learning certainly isn’t as far developed and wide spread as casual deep learning, butI think a lot of money will be invested in Reinforcement Learning once the first jobs will be automated. This holds especially true if the automated jobs are white collar jobs.