The Gegelati Framework for Efficient and Reproducible Solutions with Tangled Program Graphs

In the last decades, the need to solve complex tasks using machine learning techniques has grown significantly. While deep learningdeep learningalgorithms often achieve state-of-the-art performance, they come with high computational cost. Other methods, such as genetic programming, are able to achieve reasonable results with much lower complexity. Among the genetic programming approaches, the TPG algorithm stands out, particularly in reinforcement learningreinforcement learningapplications. In this work, we review the Gegelati framework, a framework for learning complex tasks using TPG. This framework allows efficient training with deterministicdeterministicparallelism and portability. It also supports fast inference through automatic C code generationcode generation. Gegelati demonstrates strong performance in classical reinforcement learningreinforcement learningtasks involving single discrete actions and is even able to handle applications with multiple continuous actions. Additionally, Gegelati shows promise in classificationclassificationtasks, achieving some results comparable to deep learningdeep learningmethods.