Reinforcement learning with bootstrapped value function randomization

Stanford researchers have developed a new algorithm for reinforcement learning, which can learn to take good actions with potentially long term consequences in a general unknown complex system. Unlike previous approaches, this method is able to combine complex nonlinear machine learning techniques (such as deep neural networks) with efficient experimentation in a computationally tractable manner. This method balances exploration with exploitation in way that can lead to exponential improvements over current state of the art. The system incorporates a machine learning model that receives observations of the environment and takes actions to optimize cumulative rewards. The system balances the needs for exploration with exploitation through the use of randomized value function estimates to incentivize policies that are poorly understood. This provides a balanced algorithm for automating decision and learning systems that can act and learn efficiently in terms of data, computation, and observed performance.

Visualizing Uncertainty: