Paul Templier

Inspired by biological growth, we propose SMOL to dynamically modulate robot strength during learning. This curriculum helps agents leverage advantageous physics early on, improving both performance and behavior diversity.

Quality with Just Enough Diversity (JEDi) uses behavior information from Quality-Diversity to improve the search capabilities of ES, by learning and focusing on interesting behaviors. (Follow-up work under review).

Directly optimizing the weights of a neural network with evolution can get expensive, especially with methods like XNES or CMAES. We introduce GENE, a new method to encode neural networks as genomes which keeps the performance of direct encoding while reducing the size of the genome by an order of magnitude.

While training an artificial agent for complex tasks like driving a car, mastering a video game, or controlling plasma in a nuclear fusion reactor, innovations can lead to intelligent behavior. In such scenarios, a promising approach is to mimick the natural world’s evolutionary process, which has honed the problem-solving capabilities of animal brains. Evolutionary Neural Policy Search (ENPS) draws inspiration from this concept. It creates a diverse population of “brains” represented by neural networks, allowing the system to “evolve” by selectively combining and mutating successful individuals. This thesis delves into the core components of ENPS and their intricate interplay. By analyzing the structures of ENPS, the goal is to design novel policy search methods that enhance these components, ultimately leading to the development of more efficient and effective learning algorithms for complex tasks.