GECCO '21: Proceedings of the Genetic and Evolutionary Computation ConferenceFull Citation in the ACM Digital Library
The diversity between individual learners in an ensemble is known to influence its performance. However, there is no standard agreement on how diversity should be defined, and thus how to exploit it to construct a high-performing classifier. We propose two new behavioural diversity metrics based on the divergence of errors between models. Following a neuroevolution approach, these metrics are then used to guide a novelty search algorithm to search a space of neural architectures and discover behaviourally diverse classifiers, iteratively adding the models with high diversity score to an ensemble. The parameters of each ANN are tuned individually with a standard gradient descent procedure. We test our approach on three benchmark datasets from Computer Vision --- CIFAR-10, CIFAR-100, and SVHN --- and find that the ensembles generated significantly outperform ensembles created without explicitly searching for diversity and that the error diversity metrics we propose lead to better results than others in the literature. We conclude that our empirical results signpost an improved approach to promoting diversity in ensemble learning, identifying what sort of diversity is most relevant and proposing an algorithm that explicitly searches for it without selecting for accuracy.
We demonstrate the training of Spiking Neural Networks (SNN) in a novel multi-agent Evolutionary Robotics (ER) framework inspired by competitive evolutionary environments in nature. The topology of a SNN along with morphological parameters of the bot it controls in the ER environment is together treated as a phenotype. Rules of the framework select certain bots and their SNNs for reproduction and others for elimination based on their efficacy in capturing food in a competitive environment. While the bots and their SNNs are not explicitly trained to survive or reproduce using loss functions, these drives emerge implicitly as they evolve to hunt food and survive. Their efficiency in capturing food exhibits the evolutionary signature of punctuated equilibrium. We use this signature to compare the performances of two evolutionary inheritance algorithms on the phenotypes, Mutation and Crossover with Mutation, using ensembles of 100 experiments for each algorithm. We find that Crossover with Mutation promotes 40% faster learning in the SNN than mere Mutation with a statistically significant margin.
Quality-Diversity optimization algorithms such as MAP-Elites, aim to generate collections of both diverse and high-performing solutions to an optimization problem. MAP-Elites has shown promising results in a variety of applications. In particular in evolutionary robotics tasks targeting the generation of behavioral repertoires that highlight the versatility of robots. However, for most robotics applications MAP-Elites is limited to using simple open-loop or low-dimensional controllers. Here we present Policy Gradient Assisted MAP-Elites (PGA-MAP-Elites), a novel algorithm that enables MAP-Elites to efficiently evolve large neural network controllers by introducing a gradient-based variation operator inspired by Deep Reinforcement Learning. This operator leverages gradient estimates obtained from a critic neural network to rapidly find higher-performing solutions and is paired with a traditional genetic variation to maintain a divergent search behavior. The synergy of these operators makes PGA-MAP-Elites an efficient yet powerful algorithm for finding diverse and high-performing behaviors. We evaluate our method on four different tasks for building behavioral repertoires that use deep neural network controllers. The results show that PGA-MAP-Elites significantly improves the quality of the generated repertoires compared to existing methods.
Neural Architecture Search (NAS) is the name given to a set of methods designed to automatically configure the layout of neural networks. Their success on Convolutional Neural Networks inspired its use on optimizing other types of neural network architectures, including Graph Neural Networks (GNNs). GNNs have been extensively applied over several collections of real-world data, achieving state-of-the-art results in tasks such as circuit design, molecular structure generation and anomaly detection. Many GNN models have been recently proposed, and choosing the best model for each problem has become a cumbersome and error-prone task. Aiming to alleviate this problem, recent works have proposed strategies for applying NAS to GNN models. However, different search methods converge relatively fast in the search for a good architecture, which raises questions about the structure of the problem. In this work we use Fitness Landscape Analysis (FLA) measures to characterize the search space explored by NAS methods for GNNs. We sample almost 90k different architectures that cover most of the fitness range, and represent them using both a one-hot encoding and an embedding representation. Results of the fitness distance correlation and dispersion metrics show the fitness landscape is easy to be explored, and presents low neutrality.
Neural networks with temporal characteristics such as asynchronous spiking have made much progress towards biologically plausible artificial intelligence. However, genetic approaches for evolving network structures in this area are still relatively unexplored. In this paper, we examine a specific variant of time-dependent spiking neural networks (NN) in which the spatial and temporal relationships between neurons affect output. First, we built and customized a standard NN implementation to more closely model the time-delay characteristics of biological neurons. Next, we tested this with simulated tasks such as food foraging and image recognition, demonstrating success in multiple domains. We then developed a genetic representation for the network that allows for both scalable network size and compatibility with genetic crossover operations. Finally, we analyzed the effects of genetic crossover algorithms compared to random mutations on the food foraging task. Results showed that crossover operations based on node usage converge on a local maximum more quickly than random mutations, but suffer from genetic defects that reduce overall population performance.
Evolving and merging hebbian learning rules: increasing generalization by decreasing the number of rules
Generalization to out-of-distribution (OOD) circumstances after training remains a challenge for artificial agents. To improve the robustness displayed by plastic Hebbian neural networks, we evolve a set of Hebbian learning rules, where multiple connections are assigned to a single rule. Inspired by the biological phenomenon of the genomic bottleneck, we show that by allowing multiple connections in the network to share the same local learning rule, it is possible to drastically reduce the number of trainable parameters, while obtaining a more robust agent. During evolution, by iteratively using simple K-Means clustering to combine rules, our Evolve & Merge approach is able to reduce the number of trainable parameters from 61,440 to 1,920, while at the same time improving robustness, all without increasing the number of generations used. While optimization of the agents is done on a standard quadruped robot morphology, we evaluate the agents' performances on slight morphology modifications in a total of 30 unseen morphologies. Our results add to the discussion on generalization, overfitting and OOD adaptation. To create agents that can adapt to a wider array of unexpected situations, Hebbian learning combined with a regularising "genomic bottleneck" could be a promising research direction.
Neuroevolution is an alternative to gradient-based optimisation that has the potential to avoid local minima and allows parallelisation. The main limiting factor is that usually it does not scale well with parameter space dimensionality. Inspired by recent work examining neural network intrinsic dimension and loss landscapes, we hypothesise that there exists a low-dimensional manifold, embedded in the policy network parameter space, around which a high-density of diverse and useful policies are located. This paper proposes a novel method for diversity-based policy search via Neuroevolution, that leverages learned representations of the policy network parameters, by performing policy search in this learned representation space. Our method relies on the Quality-Diversity (QD) framework which provides a principled approach to policy search, and maintains a collection of diverse policies, used as a dataset for learning policy representations. Further, we use the Jacobian of the inverse-mapping function to guide the search in the representation space. This ensures that the generated samples remain in the high-density regions, after mapping back to the original space. Finally, we evaluate our contributions on four continuous-control tasks in simulated environments, and compare to diversity-based baselines.
Previous evolution based architecture search require high computational resources resulting in large search time. In this work, we propose a novel way of applying a simple genetic algorithm to the neural architecture search problem called EvNAS (Evolving Neural Architecture using One Shot Model) which reduces the search time significantly while still achieving better result than previous evolution based methods. The architectures are represented by architecture parameter of one shot model which results in the weight sharing among the given population of architectures and also weight inheritance from one generation to the next generation of architectures. We use the accuracy of partially trained architecture on validation data as a prediction of its fitness to reduce the search time. We also propose a decoding technique for the architecture parameter which is used to divert majority of the gradient information towards the given architecture and is also used for improving the fitness prediction of the given architecture from the one shot model during the search process. EvNAS searches for architecture on CIFAR-10 for 3.83 GPU day on a single GPU with top-1 test error 2.47%, which is then transferred to CIFAR-100 and ImageNet achieving top-1 error 16.37% and top-5 error 7.4% respectively.
A major limitation to the optimization of artificial neural networks (ANN) with evolutionary methods lies in the high dimensionality of the search space, the number of weights growing quadratically with the size of the network. This leads to expensive training costs, especially in evolution strategies which rely on matrices whose sizes grow with the number of genes. We introduce a geometric encoding for neural network evolution (GENE) as a representation of ANN parameters in a smaller space that scales linearly with the number of neurons, allowing for efficient parameter search. Each neuron of the network is encoded as a point in a latent space and the weight of a connection between two neurons is computed as the distance between them. The coordinates of all neurons are then optimized with evolution strategies in a reduced search space while not limiting network fitness and possibly improving search.