A collaboration born from the Mitacs Globalink Research Internship program at St. Francis Xavier University has produced new research addressing one of the most persistent problems in modern artificial intelligence: how to teach autonomous agents what we actually want them to do.
StFX Department of Computer Science faculty Dr. Jean-Alexis Delamer and Dr. James Alexander Hughes, together with Kamalesh Kumar, a Mitacs Globalink student who joined StFX from the University of Massachusetts Amherst to work with the team, have developed a new framework called RevolveRL. Their paper, "Breaking Free from Hand-Crafted Rewards: A Genetic Programming Framework for End-Goal-Driven Reinforcement Learning," introduces a system that automatically discovers effective reward functions for reinforcement learning (RL) agents rather than requiring humans to design them by hand.
The paper has been submitted to IEEE Congress on Evolutionary Computation 2026.
Reinforcement learning is the technology behind many recent advances in AI, from robotics to game-playing systems, says Dr. Delamer. “It works by rewarding an agent for good behaviour so that it learns to make better decisions over time. But designing those reward signals is notoriously difficult. Poorly chosen rewards can lead agents to exploit shortcuts, behave unpredictably, or simply fail to learn at all. RevolveRL sidesteps this problem by using genetic programming, an evolutionary computation technique that mimics natural selection to evolve reward functions directly aligned with the true goal of a task,” he says.
The team evaluated RevolveRL across a wide range of benchmarks, from complex physics-based control tasks in the MuJoCo simulator. The evolved rewards consistently produced faster learning, more stable training, and better final performance than manually engineered baselines and several state-of-the-art reward-shaping methods, Dr. Delamer says.
“The project also illustrates the value of Mitacs Globalink at StFX: bringing talented international students into hands-on research collaborations with StFX faculty and producing high-impact results.”
The research was enabled in part by the Digital Research Alliance of Canada.
