Efficient Reinforcement Learning Through Evolving Neural Network Topologies

By Kenneth O. Stanley, Risto Miikkulainen. 2002

282 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells (2018, background): “Traditionally, architecture search methods have been relying upon evolutionary strategies [2, 39, 40], where a population of networks (oftentimes together with their weights) is continuously mutated, and less promising networks are being discarded.”
A Taxonomy for Artificial Embryogeny (2003, background): “Stanley and Miikkulainen [75, 76] also showed that historical markings can be used to speciate the population, separating incompatible organisms into different niches.”
Evolving a Roving Eye for Go (2004, methodology): “on the benchmark double pole balancing task [5,13].”
Reinforcement Learning in Continuous State and Action Spaces (2012, background): “However, it is also possible to change the structure of the function during learning (e.g., Stanley and Miikkulainen, 2002; Taylor et al, 2006; Whiteson and Stone, 2006; Buşoniu et al, 2010).”
Cooperation and communication in multiagent deep reinforcement learning (2016, background): “to optimize either the topology or parameters of the networks, sometimes in combination with backpropagation (Whiteson, 2007), and sometimes alone (Stanley and Miikkulainen, 2002).”
Learning Time/Memory-Efficient Deep Architectures with Budgeted Super Networks (2017, methodology): “Different algorithms have been proposed including evolutionary methods [27, 20, 23] or reinforcement learning-based approaches [32].”
Developmental Plasticity in Cartesian Genetic Programming based Neural Networks (2016, methodology): “Method SinglePole DoublePole CNE 352 22100 SANE 302 12600 ESP 289 3800 NEAT 743 3600 CoSyNE 98 954 FCGPANN 21 77 PCGPANN 104 1169 Evolution (SANE)(Moriarty, 1997), Enforced SubPopulation (ESP) (Gomez and Miikkulainen, 1999), Neuro Evolution of Augmenting Topologies (NEAT) (Stanley and Miikkulainen, ”
Coevolution of neural networks using a layered pareto archive (2006, background): “In particular, it has been reported to require a record low number of evaluations in the double pole balancing benchmark [18].”
The road to everywhere : Evolution, complexity and progress in natural and artificial systems (2008, background): “[171] This tagging idea does not only provide a way to control crossover; it also allows one to compare two networks and to determine their degree of “parenthood”, i.”
Picbreeder: A Case Study in Collaborative Evolutionary Exploration of Design Space (2011, methodology): “Through the compounding effect of branching, and the ability of the underlying NeuroEvolution of Augmenting Topologies (NEAT) algorithm (Stanley and Miikkulainen, 2002b, 2004a) to increase the images’ complexity, users succeed in collaboratively searching the space of possible images.”
Evolving adaptive neural networks with and without adaptive synapses (2003, methodology): “This approach is highly effective: NEAT outperforms other neuroevolution (NE) methods, e.g. on the benchmark double pole balancing task (Stanley and Miikkulainen 2002c,d).”
Real-time neuroevolution in the NERO video game (2005, background): “, on the benchmark double pole balancing task [4], [76].”
Gaussian Processes for Regression and Optimisation (2007, methodology): “Stanley and Miikkulainen [76, 75, 74] introduced “Neuroevolution of Augmenting Topologies” (NEAT), and applied it to a number of pole balancing tasks, including the double pole balancing with velocity information task, which is the same as the double pole balancing problem as presented above.”
Policy Optimization by Genetic Distillation (2017, methodology): “NEAT has also been successfully applied for policy optimization (Stanley & Miikkulainen, 2002a).”
Evolving Normalization-Activation Layers (2020, background): “, [20,25–50], but has a very different goal.”