DyLam: A Dynamic Reward Weighting Framework for Reinforcement Learning Algorithms

Creating a Reinforcement Learning (RL) training environment is a known difficulty in the field. When the reward is a composition of different signals, defining the weights for each signal represents the learning curricula the agent will follow during training. The process of trying new weights endures until the agent reaches the objective of the environment. We present DyLam, a robust automated self-curriculum learning framework for RL algorithms. By requiring only an estimate of the theoretical maximum and minimum of each reward component, DyLam can adjust its weights dynamically during training, depending on which signal needs to be optimized at each training stage. We show experimentally the robustness of this method compared to the state-of-the-art in the Lunar Lander discrete control benchmark context. CCS CONCEPTS • Computing methodologies → Reinforcement learning.