NeSSy: Neuro-symbolic Synthesis via Reinforcement Learning

Research Project, 2023 – 2026

The aim of this project is to develop a neuro-symbolic framework for synthesis of communication schemes in the context of multi-agent  reinforcement learning. Agents will collaborate to solve shared tasks, during which they will develop useful patterns of communication driven by reinforcement learning rewards. In addition, agents will also (symbolically) reflect on their results and explore revisions of their current language for subsequent interactions. Such revisions will be successful if they a) can be learnt by the other agent, and b) allow to send shorter, more informative messages. Different biases can be encoded in the reinforcement learning rewards to evaluate which factors would contribute to developing which language properties.The project contributes to central areas in AI, neuro-symbolic program synthesis and cognitive science. Our approach is unique in combining symbolic program synthesis techniques with multi-agent reinforcement learning, using the reward to steer the introduction and evaluation of novel concepts. As such we provide new perspectives and techniques in multi-agent RL and program synthesis. For cognitive science, this research brings a new paradigm into force in addressing central problems; recent research in computational learning has made it increasingly apparent that reinforcement learning offers more than just a procedure for promoting effective decision-making, but also resemble exploration strategies used by humans.