Q-Networks with Dynamically Loaded Biases for Personalization

Personalization is ever more prevalent in digital systems in various application domains. Reinforcement learning is a method often applied to adjust a system's behavior to the user's preferences, but there are a number of hurdles when applying it in this context. We propose a novel neural network architecture for reinforcement learning agents specifically tailored to support personalization - Dynamically Loaded Biases Q-Network. We test our architecture on two environments simulating a personalization task and show that it can simultaneously learn a general behavior and adjust it to different environments.