Genetic programming (GP) is a popular problem-solving and optimization technique. However, generating effective test cases for training and evaluating GP programs requires strong domain knowledge. Furthermore, GP programs often prematurely converge on local optima when given excessively difficult problems early in their training. Curriculum learning (CL) has been effective in addressing similar issues across different reinforcement learning (RL) domains, but it requires the manual generation of progressively difficult test cases as well as their careful scheduling. In this work, we leverage the domain knowledge and the strong generative abilities of large language models (LLMs) to generate effective test cases of increasing difficulties and schedule them according to various curricula. We show that by integrating a curriculum scheduler with LLM-generated test cases we can effectively train a GP agent player with environments-based curricula for a single-player game and opponent-based curricula for a multi-player game. Finally, we discuss the benefits and challenges of implementing this method for other problem domains.
Observing the ocean provides us with essential information necessary to study and understand marine ecosystem dynamics, its evolution and the impact of human activities. However, observations are sparse, limited in time and space coverage, and unevenly collected among variables. Our work aims to develop an improved deep-learning technique for predicting relationships between high-frequency and low-frequency sampled variables. Specifically, we use a larger dataset, EMODnet, and train our model for predicting nutrient concentrations and carbonate system variables (low-frequency sampled variables) starting from information such as sampling time and geolocation, temperature, salinity and oxygen (high-frequency sampled variables). Novel elements of our application include (i) the calculation of a confidence interval for prediction based on deep ensembles of neural networks, and (ii) a two-step analysis for the quality check of the input data. The proposed method proves capable of predicting the desired variables with relatively small errors, outperforming the results obtained by the current state-of-the-art models.
7 December 2022 in Swarm and Evolutionary Computation
Geometric semantic genetic programming (GSGP) is a popular form of GP where the effect of crossover and mutation can be expressed as geometric operations on a semantic space. A recent study showed that GSGP can be hybridized with a standard gradient-based optimized, Adam, commonly used in training artificial neural networks.We expand upon that work by considering more gradient-based optimizers, a deeper investigation of their parameters, how the hybridization is performed, and a more comprehensive set of benchmark problems. With the correct choice of hyperparameters, this hybridization improves the performances of GSGP and allows it to reach the same fitness values with fewer fitness evaluations.