Edgar Galván

Pramit Dutta, Ganesh Sistu, Senthil Yogamani et al.

Generating a detailed near-field perceptual model of the environment is an important and challenging problem in both self-driving vehicles and autonomous mobile robotics. A Bird Eye View (BEV) map, providing a panoptic representation, is a commonly used approach that provides a simplified 2D representation of the vehicle surroundings with accurate semantic level segmentation for many downstream tasks. Current state-of-the art approaches to generate BEV-maps employ a Convolutional Neural Network (CNN) backbone to create feature-maps which are passed through a spatial transformer to project the derived features onto the BEV coordinate frame. In this paper, we evaluate the use of vision transformers (ViT) as a backbone architecture to generate BEV maps. Our network architecture, ViT-BEVSeg, employs standard vision transformers to generate a multi-scale representation of the input image. The resulting representation is then provided as an input to a spatial transformer decoder module which outputs segmentation maps in the BEV grid. We evaluate our approach on the nuScenes dataset demonstrating a considerable improvement in the performance relative to state-of-the-art approaches.

Fergal Stapleton, Edgar Galván, Ganesh Sistu et al.

The incentive for using Evolutionary Algorithms (EAs) for the automated optimization and training of deep neural networks (DNNs), a process referred to as neuroevolution, has gained momentum in recent years. The configuration and training of these networks can be posed as optimization problems. Indeed, most of the recent works on neuroevolution have focused their attention on single-objective optimization. Moreover, from the little research that has been done at the intersection of neuroevolution and evolutionary multi-objective optimization (EMO), all the research that has been carried out has focused predominantly on the use of one type of DNN: convolutional neural networks (CNNs), using well-established standard benchmark problems such as MNIST. In this work, we make a leap in the understanding of these two areas (neuroevolution and EMO), regarded in this work as neuroevolutionary multi-objective, by using and studying a rich DNN composed of a CNN and Long-short Term Memory network. Moreover, we use a robust and challenging vehicle trajectory prediction problem. By using the well-known Non-dominated Sorting Genetic Algorithm-II, we study the effects of five different objectives, tested in categories of three, allowing us to show how these objectives have either a positive or detrimental effect in neuroevolution for trajectory prediction in autonomous vehicles.

Fergal Stapleton, Edgar Galván

Evolutionary Algorithms (EAs) play a crucial role in the architectural configuration and training of Artificial Deep Neural Networks (DNNs), a process known as neuroevolution. However, neuroevolution is hindered by its inherent computational expense, requiring multiple generations, a large population, and numerous epochs. The most computationally intensive aspect lies in evaluating the fitness function of a single candidate solution. To address this challenge, we employ Surrogate-assisted EAs (SAEAs). While a few SAEAs approaches have been proposed in neuroevolution, none have been applied to truly large DNNs due to issues like intractable information usage. In this work, drawing inspiration from Genetic Programming semantics, we use phenotypic distance vectors, outputted from DNNs, alongside Kriging Partial Least Squares (KPLS), an approach that is effective in handling these large vectors, making them suitable for search. Our proposed approach, named Neuro-Linear Genetic Programming surrogate model (NeuroLGP-SM), efficiently and accurately estimates DNN fitness without the need for complete evaluations. NeuroLGP-SM demonstrates competitive or superior results compared to 12 other methods, including NeuroLGP without SM, convolutional neural networks, support vector machines, and autoencoders. Additionally, it is worth noting that NeuroLGP-SM is 25% more energy-efficient than its NeuroLGP counterpart. This efficiency advantage adds to the overall appeal of our proposed NeuroLGP-SM in optimising the configuration of large DNNs.

Fergal Stapleton, Brendan Cody-Kenny, Edgar Galván

Evolutionary algorithms are increasingly recognised as a viable computational approach for the automated optimisation of deep neural networks (DNNs) within artificial intelligence. This method extends to the training of DNNs, an approach known as neuroevolution. However, neuroevolution is an inherently resource-intensive process, with certain studies reporting the consumption of thousands of GPU days for refining and training a single DNN network. To address the computational challenges associated with neuroevolution while still attaining good DNN accuracy, surrogate models emerge as a pragmatic solution. Despite their potential, the integration of surrogate models into neuroevolution is still in its early stages, hindered by factors such as the effective use of high-dimensional data and the representation employed in neuroevolution. In this context, we address these challenges by employing a suitable representation based on Linear Genetic Programming, denoted as NeuroLGP, and leveraging Kriging Partial Least Squares. The amalgamation of these two techniques culminates in our proposed methodology known as the NeuroLGP-Surrogate Model (NeuroLGP-SM). For comparison purposes, we also code and use a baseline approach incorporating a repair mechanism, a common practice in neuroevolution. Notably, the baseline approach surpasses the renowned VGG-16 model in accuracy. Given the computational intensity inherent in DNN operations, a singular run is typically the norm. To evaluate the efficacy of our proposed approach, we conducted 96 independent runs. Significantly, our methodologies consistently outperform the baseline, with the SM model demonstrating superior accuracy or comparable results to the NeuroLGP approach. Noteworthy is the additional advantage that the SM approach exhibits a 25% reduction in computational requirements, further emphasising its efficiency for neuroevolution.

Edgar Galván, Gavin Simpson

Monte Carlo Tree Search (MCTS) is a sampling best-first method to search for optimal decisions. The MCTS's popularity is based on its extraordinary results in the challenging two-player based game Go, a game considered much harder than Chess and that until very recently was considered infeasible for Artificial Intelligence methods. The success of MCTS depends heavily on how the tree is built and the selection process plays a fundamental role in this. One particular selection mechanism that has proved to be reliable is based on the Upper Confidence Bounds for Trees, commonly referred as UCT. The UCT attempts to nicely balance exploration and exploitation by considering the values stored in the statistical tree of the MCTS. However, some tuning of the MCTS UCT is necessary for this to work well. In this work, we use Evolutionary Algorithms (EAs) to evolve mathematical expressions with the goal to substitute the UCT mathematical expression. We compare our proposed approach, called Evolution Strategy in MCTS (ES-MCTS) against five variants of the MCTS UCT, three variants of the star-minimax family of algorithms as well as a random controller in the Game of Carcassonne. We also use a variant of our proposed EA-based controller, dubbed ES partially integrated in MCTS. We show how the ES-MCTS controller, is able to outperform all these 10 intelligent controllers, including robust MCTS UCT controllers.

Edgar Galván, Leonardo Trujillo, Fergal Stapleton

Semantics has become a key topic of research in Genetic Programming (GP). Semantics refers to the outputs (behaviour) of a GP individual when this is run on a data set. The majority of works that focus on semantic diversity in single-objective GP indicates that it is highly beneficial in evolutionary search. Surprisingly, there is minuscule research conducted in semantics in Multi-objective GP (MOGP). In this work we make a leap beyond our understanding of semantics in MOGP and propose SDO: Semantic-based Distance as an additional criteriOn. This naturally encourages semantic diversity in MOGP. To do so, we find a pivot in the less dense region of the first Pareto front (most promising front). This is then used to compute a distance between the pivot and every individual in the population. The resulting distance is then used as an additional criterion to be optimised to favour semantic diversity. We also use two other semantic-based methods as baselines, called Semantic Similarity-based Crossover and Semantic-based Crowding Distance. Furthermore, we also use the NSGA-II and the SPEA2 for comparison too. We use highly unbalanced binary classification problems and consistently show how our proposed SDO approach produces more non-dominated solutions and better diversity, leading to better statistically significant results, using the hypervolume results as evaluation measure, compared to the rest of the other four methods.

Fergal Stapleton, Edgar Galván

Semantic diversity in Genetic Programming has proved to be highly beneficial in evolutionary search. We have witnessed a surge in the number of scientific works in the area, starting first in discrete spaces and moving then to continuous spaces. The vast majority of these works, however, have focused their attention on single-objective genetic programming paradigms, with a few exceptions focusing on Evolutionary Multi-objective Optimization (EMO). The latter works have used well-known robust algorithms, including the Non-dominated Sorting Genetic Algorithm II and the Strength Pareto Evolutionary Algorithm, both heavily influenced by the notion of Pareto dominance. These inspiring works led us to make a step forward in EMO by considering Multi-objective Evolutionary Algorithms Based on Decomposition (MOEA/D). We show, for the first time, how we can promote semantic diversity in MOEA/D in Genetic Programming.

Edgar Galván

A variety of methods have been applied to the architectural configuration and learning or training of artificial deep neural networks (DNN). These methods play a crucial role in the success or failure of the DNN for most problems and applications. Evolutionary Algorithms (EAs) are gaining momentum as a computationally feasible method for the automated optimisation of DNNs. Neuroevolution is a term which describes these processes of automated configuration and training of DNNs using EAs. However, the automatic design and/or training of these modern neural networks through evolutionary algorithms is computanalli expensive. Kimura's neutral theory of molecular evolution states that the majority of evolutionary changes at molecular level are the result of random fixation of selectively neutral mutations. A mutation from one gene to another is neutral if it does not affect the phenotype. This work discusses how neutrality, given certain conditions, can help to speed up the training/design of deep neural networks.

Edgar Galván, Fergal Stapleton

The study of semantics in Genetic Program (GP) deals with the behaviour of a program given a set of inputs and has been widely reported in helping to promote diversity in GP for a range of complex problems ultimately improving evolutionary search. The vast majority of these studies have focused their attention in single-objective GP, with just a few exceptions where Pareto-based dominance algorithms such as NSGA-II and SPEA2 have been used as frameworks to test whether highly popular semantics-based methods, such as Semantic Similarity-based Crossover (SSC), helps or hinders evolutionary search. Surprisingly it has been reported that the benefits exhibited by SSC in SOGP are not seen in Pareto-based dominance Multi-objective GP. In this work, we are interested in studying if the same carries out in Multi-objective Evolutionary Algorithms based on Decomposition (MOEA/D). By using the MNIST dataset, a well-known dataset used in the machine learning community, we show how SSC in MOEA/D promotes semantic diversity yielding better results compared to when this is not present in canonical MOEA/D.

Edgar Galván, Fergal Stapleton

Semantics in the context of Genetic Program (GP) can be understood as the behaviour of a program given a set of inputs and has been well documented in improving performance of GP for a range of diverse problems. There have been a wide variety of different methods which have incorporated semantics into single-objective GP. The study of semantics in Multi-objective (MO) GP, however, has been limited and this paper aims at tackling this issue. More specifically, we conduct a comparison of three different forms of semantics in MOGP. One semantic-based method, (i) Semantic Similarity-based Crossover (SSC), is borrowed from single-objective GP, where the method has consistently being reported beneficial in evolutionary search. We also study two other methods, dubbed (ii) Semantic-based Distance as an additional criteriOn (SDO) and (iii) Pivot Similarity SDO. We empirically and consistently show how by naturally handling semantic distance as an additional criterion to be optimised in MOGP leads to better performance when compared to canonical methods and SSC. Both semantic distance based approaches made use of a pivot, which is a reference point from the sparsest region of the search space and it was found that individuals which were both semantically similar and dissimilar to this pivot were beneficial in promoting diversity. Moreover, we also show how the semantics successfully promoted in single-objective optimisation does not necessary lead to a better performance when adopted in MOGP.

Edgar Galván, Oxana Gorshkova, Peter Mooney et al.

Multiple Artificial Intelligence (AI) methods have been proposed over recent years to create controllers to play multiple video games of different nature and complexity without revealing the specific mechanics of each of these games to the AI methods. In recent years, Evolutionary Algorithms (EAs) employing rolling horizon mechanisms have achieved extraordinary results in these type of problems. However, some limitations are present in Rolling Horizon EAs making it a grand challenge of AI. These limitations include the wasteful mechanism of creating a population and evolving it over a fraction of a second to propose an action to be executed by the game agent. Another limitation is to use a scalar value (fitness value) to direct evolutionary search instead of accounting for a mechanism that informs us how a particular agent behaves during the rolling horizon simulation. In this work, we address both of these issues. We introduce the use of a statistical tree that tackles the latter limitation. Furthermore, we tackle the former limitation by employing a mechanism that allows us to seed part of the population using Monte Carlo Tree Search, a method that has dominated multiple General Video Game AI competitions. We show how the proposed novel mechanism, called Statistical Tree-based Population Seeding, achieves better results compared to vanilla Rolling Horizon EAs in a set of 20 games, including 10 stochastic and 10 deterministic games.

Edgar Galván, Peter Mooney

A variety of methods have been applied to the architectural configuration and learning or training of artificial deep neural networks (DNN). These methods play a crucial role in the success or failure of the DNN for most problems and applications. Evolutionary Algorithms (EAs) are gaining momentum as a computationally feasible method for the automated optimisation and training of DNNs. Neuroevolution is a term which describes these processes of automated configuration and training of DNNs using EAs. While many works exist in the literature, no comprehensive surveys currently exist focusing exclusively on the strengths and limitations of using neuroevolution approaches in DNNs. Prolonged absence of such surveys can lead to a disjointed and fragmented field preventing DNNs researchers potentially adopting neuroevolutionary methods in their own research, resulting in lost opportunities for improving performance and wider application within real-world deep learning problems. This paper presents a comprehensive survey, discussion and evaluation of the state-of-the-art works on using EAs for architectural configuration and training of DNNs. Based on this survey, the paper highlights the most pertinent current issues and challenges in neuroevolution and identifies multiple promising future research directions.