TY - GEN
T1 - Evolutionary Algorithms for Segment Optimization in Vectorial GP
AU - Fleck, Philipp
AU - Silva, Sara
AU - Winkler, Stephan
AU - Vanneschi, Leonardo
AU - Kommenda, Michael
AU - Affenzeller, Michael
N1 - Publisher Copyright:
© 2023 Copyright held by the owner/author(s).
PY - 2023/7/15
Y1 - 2023/7/15
N2 - Vectorial Genetic Programming (Vec-GP) extends regular GP by allowing vectorial input features (e.g. time series data), while retaining the expressiveness and interpretability of regular GP. The availability of raw vectorial data during training, not only enables Vec-GP to select appropriate aggregation functions itself, but also allows Vec-GP to extract segments from vectors prior to aggregation (like windows for time series data). This is a critical factor in many machine learning applications, as vectors can be very long and only small segments may be relevant. However, allowing aggregation over segments within GP models makes the training more complicated. We explore the use of common evolutionary algorithms to help GP identify appropriate segments, which we analyze using a simplified problem that focuses on optimizing aggregation segments on fixed data. Since the studied algorithms are to be used in GP for local optimization (e.g. as mutation operator), we evaluate not only the quality of the solutions, but also take into account the convergence speed and anytime performance. Among the evaluated algorithms, CMA-ES, PSO and ALPS show the most promising results, which would be prime candidates for evaluation within GP.
AB - Vectorial Genetic Programming (Vec-GP) extends regular GP by allowing vectorial input features (e.g. time series data), while retaining the expressiveness and interpretability of regular GP. The availability of raw vectorial data during training, not only enables Vec-GP to select appropriate aggregation functions itself, but also allows Vec-GP to extract segments from vectors prior to aggregation (like windows for time series data). This is a critical factor in many machine learning applications, as vectors can be very long and only small segments may be relevant. However, allowing aggregation over segments within GP models makes the training more complicated. We explore the use of common evolutionary algorithms to help GP identify appropriate segments, which we analyze using a simplified problem that focuses on optimizing aggregation segments on fixed data. Since the studied algorithms are to be used in GP for local optimization (e.g. as mutation operator), we evaluate not only the quality of the solutions, but also take into account the convergence speed and anytime performance. Among the evaluated algorithms, CMA-ES, PSO and ALPS show the most promising results, which would be prime candidates for evaluation within GP.
KW - evolutionary algorithms
KW - genetic programming
KW - symbolic regression
KW - vectorial
UR - http://www.scopus.com/inward/record.url?scp=85169025465&partnerID=8YFLogxK
U2 - 10.1145/3583133.3590668
DO - 10.1145/3583133.3590668
M3 - Conference contribution
AN - SCOPUS:85169025465
T3 - GECCO 2023 Companion - Proceedings of the 2023 Genetic and Evolutionary Computation Conference Companion
SP - 439
EP - 442
BT - GECCO 2023 Companion - Proceedings of the 2023 Genetic and Evolutionary Computation Conference Companion
PB - Association for Computing Machinery, Inc
T2 - 2023 Genetic and Evolutionary Computation Conference Companion, GECCO 2023 Companion
Y2 - 15 July 2023 through 19 July 2023
ER -