Tangled Program Graph for Reinforcement Learning Tasks.

The Tangled Program Graph (TPG) is a genetic programming (GP) framework that addresses high-dimensional Multi-Task Teinforcement Learning (MTRL) through emergent modularity. A bottom-up process is assumed in which multiple programs self-organise into collective decision-making entities, or teams, which then further develop into multi-team policy graphs.
The framework has been applied to many RL applications, ranging from discrete tasks such as Atari games to continuous control in the MuJoCo suite.
To use TPG efficiently, we have implemented a C++ library called Gegelati that allow efficient deterministic and portable training. We are currently working on a binding python to use the library from python with C++ efficiency.
Recently, to learn continuous control tasks involving multiple actions, such as those in the MuJoCo suite, we have proposed MAPLE: an algorithm inspired by TPG that learns better policies with smaller models, surpassing the state of the art in GP.
More recently, we have extended a hybrid version of MAPLE and TPG to learn a new MTRL benchmark on the Half Cheetah Mujoco environment, in which the Half Cheetah must avoid obstacles.
Finally, we are exploring various approaches to enhance the performance of MAPLE (and subsequently TPG) by leveraging Quality Diversity algorithms such as MapElites, enabling MAPLE to learn the hypercomplex Humanoid environment of Mujoco with realistic probabilities.” Could someone add that to the website? I will send an email later