Deep Learning Essentials

Deep reinforcement learning

As explained in previous sections, there are few core components of any RL-based system. A value function v (s,θ) or a Q-function Q (s, a, θ), and a policy function π (a|s, θ), which could be model-free or model based. The wide scale applicability of any RL-based system depends on how good these estimations are. In practice, existing Q-learning systems suffer from multiple drawbacks:

• Curse of dimensionality: If you were to apply a Q-learning-based technique to a high-dimensional RL setting, such as predicting the next joystick movement based on current pixel values of screen images of a game. A 32 x 32 sized image with boolean pixel values will lead to a total number of 2¹⁰²⁴ states. Q-learning will need a large ...