Reinforcement Learning algorithms train an agent from the reward signal it obtains. Thus, a naïve agent relies on exploration to collect its first rewards. Random actions can proved to be efficient in some setting, like illustrated in Maze DQN . But, in environments where rewards are sparse, this exploration can prove to be overly long and hazardous. This is where curiosity comes at the rescue !
How can we make the most of curiosity to explore reward-sparse environments ?

Reinforcement learning techniques shows great potential for game-based AI but fails to scale on real-world applications. Indeed, state space and action space become continuous and, thus, prevent any tabular-based learning. Therefore, reinforcement learning agents require to be piloted by a learning decision function thus falling back on deep learning.
How reinforcement learning and deep learning team up to give birth to an efficient agent ?

DeepMind’s algorithm AlphaGo amazed the whole world by beating 4-1 Go champion Lee Sedol in 2016. This complex algorithm learned to play Go by crunching thousands Go grandmasters plays. AlphaZero, DeepMind’s latest breakthrough in the field is far more impressive. It easily outperformed AlphaGo by solely learning on its own.
How was that made possible and what mathematical mechanisms lie underneath?