NEAT in increasingly non-linear control situations

Evolution of neural networks, as implemented in NEAT, has proven itself successful on a variety of low-level control problems such as pole balancing and vehicle control. Nonetheless, high-level control problems still seem to trouble neuroevolution approaches. This paper presents such a complex task and explores how different aspects of problem difficulty have varying strong influences on NEAT’s performance. Based on these findings, the question is discussed why certain problem domains are less beneficial for neuroevolution approaches’ performance, which may provide useful insights into how to design the next generation of neuroevolution algorithms. Categories and Subject Descriptors I.2.6 [Artificial Intelligence]: Learning—Connectionism and neural nets; I.2.9 [Artificial Intelligence]: Robotics— Kinematics and dynamics General Terms Algorithms, Performance, Experimentation Keywords Neuroevolution, Adaptive control, Dynamic control, NEAT

• Type: Conference
• Year: 2009
• Where: GECCO
• Authors: Matthias J. Linhardt, Martin V. Butz
• Keywords: Control Problems, GECCO 2009, Low-level Control Problems, Neuroevolution Approaches, Optimization