Grammar-Guided Genetic Programming is already outperforming humans at creating efficient transmission schedulers for large heterogeneous communications networks. We have previously proposed a multi-level grammar approach which achieved significantly better results than the canonical Grammar-Guided Genetic Programming approach. Initially, a restricted `small' grammar is utilised in order to discover suitable structures. A full grammar is then adopted after this initial phase. Hence, evolution can focus on maximising performance, by fine-tuning the well-structured models. In this work, we propose to use a hierarchical approach by employing multiple small grammars instead of a unique small grammar at the lower level, in conjunction with the full grammar at the upper level. To use multiple small grammars while maintaining the same computational budget, we have to use either (i) reduce the number of generations, or (ii) reduce the size of the population for the evolution with each of the small grammars. In this work, we confirm that the hierarchical grammar approach using the division of number of generations strategy achieves significantly better results than the multi-level approach, but requires defining an ideal number of small grammars to achieve the best performance. We also show that the hierarchical grammar approach using the division of population size strategy achieves significantly better results than the multi-level approach. However the division of population size strategy is less sensitive to the number of small grammars.

We present a novel approach to generating seasonal training plans for elite athletes using the grammatical evolution approach to genetic programming. A grammatical encoding of a team sport training plan dictates the plan structure. The quality of the training plan is calculated using the widely adopted fitness-fatigue model, which in this study incorporates four performance metrics, namely distance covered at low to medium speed, distance covered at high speed, distance covered accelerating, and distance covered decelerating. We compare performance of the evolved training plans to a control setup which generates plans using a pseudo-random search process, and baseline against the training plan adopted by an elite team of Gaelic Football Players. Significant potential performance gains are achieved over the control setup and baseline elite team plan.

Evolutionary computation is used to automatically evolve small cell schedulers on a realistic simulation of a 4G-LTE heterogeneous cellular network. Evolved schedulers are then further augmented by human design to improve robustness. Extensive analysis of evolved solutions and their performance across a wide range of metrics reveals evolution has uncovered a new human-competitive scheduling technique which generalises well across cells of varying sizes. Furthermore, evolved methods are shown to conform to accepted scheduling frameworks without the evolutionary process being explicitly told the form of the desired solution. Evolved solutions are shown to out-perform a human-engineered state-of-the-art benchmark by up to 50%. Finally, the approach is shown to be flexible in that tailored algorithms can be evolved for specific scenarios and corner cases, allowing network operators to create unique algorithms for different deployments, and to postpone the need for costly hardware upgrades.

Traditional single-tiered wireless communications networks cannot scale to satisfy exponentially rising demand. Operators are increasing capacity by densifying their existing macro cell deployments with co-channel small cells. However, cross-tier interference and load balancing issues present new optimization challenges in channel sharing heterogeneous networks (HetNets). One-size-fits-all heuristics for allocating resources are highly suboptimal, but designingad hoccontrollers requires significant human expertise and manual fine-tuning. In this paper, a unified, flexible, and fully automated approach for end-to-end optimization in multi-layer HetNets is presented. A hill climbing algorithm is developed for reconfiguring cells in real time in order to track dynamic traffic patterns. Schedulers for allocating spectrum to user equipment are automatically synthesized using grammar-based genetic programming. The proposed methods for configuring the HetNet and scheduling in the time–frequency domain can addressad hocobjective functions. Thus, the operator can flexibly tune the tradeoff between peak rates and fairness. Far cell edge downlink rates are increased by up to 250% compared with non-adaptive baselines. Alternatively, peak rates are increased by up to 340%. The experiments illustrate the utility and future potential of natural computing techniques in software-defined wireless communications networks.

Program synthesis is a complex problem domain tackled by many communities via different methods. In the last few years, a lot of progress has been made with Genetic Programming (GP) on solving a variety of general program synthesis problems for which a benchmark suite has been introduced. While Genetic Programming is capable of finding correct solutions for many problems contained in a general program synthesis problems benchmark suite, the actual success rate per problem is low in most cases. In this paper, we analyse certain aspects of the benchmark suite and the computational effort required to solve its problems. A subset of problems on which GP performs poorly is identified. This subset is analysed to find measures to increase success rates for similar problems. The paper concludes with suggestions to refine performance on program synthesis problems.

Small Cells are being deployed alongside pre-existing Macro Cells in order to satisfy demand during the current era of exponential growth in mobile traffic. Heterogeneous networks are economical because both cell tiers share the same scarce and expensive spectrum. However, customers at cell edges experience severe cross-tier interference in channel sharing Het-Nets, resulting in poor service quality. Techniques for improving fairness globally have been developed in previous works. In this paper, a novel method for service differentiation at the level of individual customers is proposed. The proposed algorithm redistributes spectrum on a millisecond timescale, so that premium customers experience minimum downlink rates exceeding a target threshold. System level simulations indicate that downlink rate targets of at least 1 [Mbps] are always satisfied under the proposed scheme. By contrast, naive scheduling achieves the 1 [Mbps] target only 83% of the time. Quality of service can be improved for premium customers without significantly impacting global fairness metrics. Flexible service differentiation will be key to effectively monetizing the next generation of 5G wireless communications networks.

The use of semantic information in genetic programming operators has shown major improvements in recent years, especially in the regression and boolean domain. As semantic information is domain specific, using it in other areas poses certain problems. Semantic operators require being adapted for the problem domain they are applied to. An attempt to create a semantic crossover for program synthesis has been made with rather limited success, but the results have provided insights about using semantics in program synthesis. Based on this initial attempt, this paper presents an improved version of semantic operators for program synthesis, which contains a small but significant change to the overall functionality, as well as a novel measure for the comparison of the semantics of subtrees. The results show that the improved semantic crossover is superior to the previous semantic operator in the program synthesis domain.

Drawing on a rich literature concerning social learning in animals, this paper presents a variation of Grammatical Evolution (GE) which incorporates one of the most powerful forms of social learning, namely imitation learning. This replaces the traditional method of `communication' between individuals in GE - crossover - which is drawn from an evolutionary metaphor. The paper provides an introduction to social learning, describes the proposed variant of GE, and tests on a series of benchmark symbolic regression problems. The results obtained are encouraging, being very competitive when compared with canonical GE. It is noted that the literature on social learning provides a number of useful meta-frameworks which can be used in the design of new search algorithms and to allow us to better understand the strengths and weaknesses of existing algorithms. Future work is indicated in this area.

This paper presents gems, a novel method to accelerate fitness improvement in Evolutionary Algorithms (EAs). The paper develops the models, describes an experimental implementation, comments on characteristics of problem-domains that indicate where gems may be used, and suggests an explanation of the observed behavior. Experimental results show that gems accelerate the rate of fitness increase, and that the larger the problem instance, the larger the benefit. Runtime analysis shows that the method's fitness boost far outweighs its performance costs.

Genetic Algorithms (GAs) have been shown to be a very effective optimisation tool on a wide variety of problems. However, they are not without their drawbacks. GAs require time to run, and evolve a bespoke solution to the desired problem in real time. This requirement can prove to be prohibitive in a high-frequency dynamic environment where on-line training time is limited. Neural Networks (NNs) on the other hand can be trained at length off-line, before being deployed on-line, allowing for fast generation of solutions on demand. This study presents a hybrid approach to time-frame scheduling in a high frequency domain. A GA approach is used to generate a dataset of optimised human-competitive solutions. Deep Learning is then deployed to extract the underlying model within the GA, enabling fast optimisation on unseen data. This hybrid approach allows for NNs to generate GA-quality schedules on-line, almost 100 times faster than running the GA.

Initialisation in Grammatical Evolution (GE) is a topic that remains open to debate on many fronts. The literature falls between two mainstay approaches: random and sensible initialisation. These methods are not without their drawbacks with the type of trees generated. This paper tackles this problem by extending these traditional operators to incorporate position independence in the initialisation process in GE. This new approach to initialisation is shown to provide a viable alternative to the commonly used approaches, whilst avoiding the common pitfalls of traditional approaches to initialisation.

We explore the application of grammar-based Genetic Programming, specifically Grammatical Evolution, to the problem of modeling the outcome of Six Nations Rugby matches. A series of grammars are developed in attempts to generate different forms of predictive rules, which might be useful in pre-match and mid-match scenarios. A number of interesting models are generated and their utility discussed.

A training protocol for learning deep neural networks, called greedy layer-wise training, is applied to the evolution of a hierarchical, feed-forward Genetic Programming based system for feature construction and object recognition. Results on a popular handwritten digit recognition benchmark clearly demonstrate that two layers of feature transformations improves generalisation compared to a single layer. In addition, we show that the proposed system outperforms several standard Genetic Programming systems, which are based on hand-designed features, and use different program representations and fitness functions.

We present an investigation into the genotype-phenotype map in Position Independent Grammatical Evolution (πGE). Previous studies have shown πGE to exhibit a performance increase over standard Grammatical Evolution (GE). The only difference between the two approaches is in how the genotype-phenotype mapping process is performed. GE uses a leftmost non terminal expansion, while πGE evolves the order of mapping as well as the content. In this study, we use the idea of focused search to examine which aspect of the πGE mapping process provides the lift in performance over standard GE by applying our approaches to three benchmark problems taken from specialised literature. We examined the traditional πGE approach and compared it to two setups which examined the extremes of mapping order search and content search, and against setups with varying ratios of content and order search. In all of these tests a purely content focused πGE was shown to exhibit a performance gain over the other setups.

In this work, we examine the capabilities of two forms of mappings by means of Grammatical Evolution (GE) to successfully generate controllers by combining high-level functions in a dynamic environment. In this work we adopted the Ms. Pac-Man game as a benchmark test bed. We show that the standard GE mapping and Position Independent GE (πGE) mapping achieve similar performance in terms of maximising the score. We also show that the controllers produced by both approaches have an overall better performance in terms of maximising the score compared to a hand-coded agent. There are, however, significant differences in the controllers produced by these two approaches: standard GE produces more controllers with invalid code, whereas the opposite is seen with πGE.