Home Data Deep reinforcement learning – trick or treat?

Deep reinforcement learning – trick or treat?

Halloween in The Future
2 min read

Deep Reinforcement Learning (Deep RL) is the new buzzword in the machine learning world. Deep RL is an approach which combines reinforcement learning and deep learning in order to achieve human-level performance. It brings together the self-learning approach to learn successful strategies that lead to the greatest long-term rewards and allows the agents to construct and learn their own knowledge directly from raw inputs.

With the fusion of these two approaches, we saw the introduction of many algorithms, starting with DeepMind’s Deep Q Network (DQN). It is a deep variant of the Q-learning algorithm. This algorithm reached human-level performance in playing Atari games. Combining Q-learning with reasonably sized neural networks and some optimization tricks, you can achieve human or superhuman performance in several Atari games.

Learn Programming & Development with a Packt Subscription

Deep RL resulted in one of the notable advancements in the game of AlphaGo.The AI agent by DeepMind was able to beat the human world champions Lee Sedol (4-1) and Fan Hui (5-0). DeepMind then further released advanced versions of their Agent called AlphaGO Zero and AlphaZero. Many recent works from the researchers at UC Berkeley have shown how both reinforcement learning and deep reinforcement learning have enabled the control of complex robots, both for locomotion and navigation.

Despite these successes, it is quite difficult to find cases where deep RL has added any practical real-world value. The current status is that it is still a research topic. One of its limitations is that it assumes the existence of a reward function, which is either given or is hand-tuned offline. To get the desired results, your reward function must capture exactly what you want. RL has an annoying tendency to overfit to your reward, resulting in things you haven’t expected. This is the reason why Atari is a benchmark, as it is not only easy to get a lot of samples, but the goal is fairly straightforward i.e to maximize score.

With so many researchers working towards introducing improved Deep RL algorithms, it surely is a treat.

Read Next

AlphaZero: The genesis of machine intuition

DeepMind open sources TRFL, a new library of reinforcement learning building blocks

Understanding Deep Reinforcement Learning by understanding the Markov Decision Process [Tutorial]