Show HN: Watch a neural net discover molecules by arguing with itself (randman444.github.io)

A new AI project showcases a neural network capable of discovering novel molecules through an innovative self-argumentation approach. The system employs a reinforcement learning (RL) framework where two identical clones, one cautious and one curious, generate SMILES strings for molecules in a competitive setting. Each clone trains side-by-side and assesses its performance through a metric called epiplexity, which reflects the learning improvement of the critic. The clone that demonstrates superior learning not only wins but also passes its improved weights to the next round, allowing the network to adapt and refine its search strategies autonomously. This method is particularly significant for the AI/ML community as it enables the neural network to self-tune its exploration and exploitation parameters without human intervention, especially in environments with sparse rewards. By doing so, the model can intelligently navigate complex search spaces for molecular design, showcasing a potential breakthrough in drug discovery. Early results demonstrate that while the dual-agent system may not outperform existing baselines in scenarios with dense rewards, it excels in sparse environments, effectively adjusting its exploration strategy to find viable candidates like Tanimoto-similar molecules to established drugs such as aspirin.