Code for "Learning from Interpretable Goals: Dynamic Weight Balancing for Multi-Objective Reinforcement Learning"

Reward engineering remains a central challenge in reinforcement learning (RL), particularly when multiple, potentially conflicting objectives must be balanced. Often, this is addressed through a weighted sum of reward components, yet this approach introduces several difficulties: extensive manual tuning of weights, lack of semantic interpretability, and limited generality of methods that rely on algorithm-specific extensions.In this work, we address these challenges by proposing a goal-based reward formulation that replaces opaque weight tuning with interpretable goal definitions. These goals are learned through a curriculum-based training scheme that eliminates the need for manual weight tuning by dynamically adapting weights while remaining agnostic to the underlying RL algorithm.We evaluate our approach on five control environments from the Farama Gymnasium and show that it achieves comparable or superior success rates to the original tasks. Moreover, our method eliminates the need for extensive reward weighting and introduces an interpretable training framework for multi-objective reinforcement learning.