Thrilled to share that our paper “On the Importance of Reward Design in Reinforcement Learning-based Dynamic Algorithm Configuration: A Case Study on OneMax with (1+(λ,λ))-GA” received the Best Paper Award at GECCO 2025 (Málaga, Spain, July 14–18), selected from a competitive pool with a 36% acceptance rate.
The work introduces an adaptive reward shifting mechanism that prevents RL agents from getting trapped in safe-but-suboptimal policies, beating theoretical baselines on OneMax-DAC while training in just 12,000 steps — orders of magnitude fewer than incumbent tuners like IRACE.