Real World Optimisation with Life-Long Learning

工业领域中出现的许多涉及可持续运营、满足需求和将成本降至最低的实际问题无法得到准确解决。元启发式优化技术已经在学术界被广泛开发来解决这类问题，并在文献中报道了许多成功的情况。然而，在对优化技术的科学研究与终端用户面临并由商业优化软件供应商解决的那些问题之间，仍然存在一个令人担忧的空白。从商业角度来看，与现实世界面临的问题相比，学术界解决的问题过于简单化，未能接受现实世界的许多限制。从科学的角度来看，研究人员还发现“商业软件中缺乏先进的元启发式技术”，这部分归因于学术界未能证明他们的解决方案适用于商业世界的需求，部分原因在于学术界未能向工业界传达他们的信息。元启发式方法的开发成本可能很高，因为它们通常需要人类专业知识来将专业知识整合到算法中，并需要启发式方法的专业知识来设计和调整算法。因此，最近的研究集中在自动算法设计和配置上，这些算法产生的优化求解器无论是在单个问题上还是在一组问题上都表现良好。该领域的一个分支是超启发式算法，它在低级启发式算法的空间上运行，寻找利用单个已知启发式算法的优点和缺点的启发式算法的组合。由此产生的算法实现成本低，需要的人工专业知识较少，在问题类中具有健壮的性能，并且可以跨问题域移植。与量身定制的方法相比，这些功能弥补了解决方案质量的一些下降，同时仍确保解决方案具有可接受的质量。然而，大多数自动化设计方法未能纳入或认识到人类的一项关键能力；人类不断从经验中学习--通过概括观察和反馈，他们能够更新其内部问题解决模型，以便不断改进它们，并适应不断变化的环境。计算求解器未能利用先前的知识，既浪费了有用的知识，又潜在地阻碍了发现好的解决方案。此外，如果领域中问题实例的特征随着时间的推移而改变，则可能需要完全重新调整解算器，或者在最坏的情况下周期性地重新设计。我们提出了一种新颖的终身学习超启发式系统，它解决了目前系统固有的缺陷：它将表现出短期学习，为单个问题产生快速有效的解决方案，同时，长期学习过程将使系统能够随着时间的推移自主适应新的问题特征。因此，它利用现有的知识，同时适应新的信息。其次，通过与两个合作者，一个商业路线软件供应商和一个林业专家的密切合作，我们的研究将直接从现实世界的问题中获得信息，考虑到真实的约束和性能标准，从而产生经济影响。优化技术的未来进步将通过问题生成器和一些反映现实世界优先事项和限制的问题套件的开发而得到促进，这些问题套件来自我们的合作者提供的实际问题数据，并与不仅反映经济驱动因素而且解决环境影响和碳排放减少的指标一起定义。将广泛传播这一信息数据库，为今后的研究提供广泛的平台。

Many practical problems arising in industrial domains concerned with operating sustainably, meeting demand and minimising costs cannot be solved exactly. Meta-heuristic optimisation techniques have been widely developed in academia to solve such problems with much success reported in the literature. However, there remains a worrying void between scientific research into optimisation techniques and those problems faced by end-users and addressed by commercial optimisation software vendors. From a commercial perspective, the problems addressed by academia are too simplistic compared to those faced in the real-world, failing to embrace many real-world constraints. From the scientific perspective, researchers have also identified a "lack of advanced metaheuristic techniques in commercial software'' which has been attributed in part to the academic community failing to demonstrate that their solutions are applicable to the needs of the commercial world, and in part to academics failing to impart their message the industrial community.Meta-heuristic approaches can be costly to develop as they generally require human expertise to integrate specialist knowledge into an algorithm, and expertise in heuristic methods to design and tune algorithms. Recent research has therefore focused on automated algorithm design and configuration which produce tuned solvers that perform well on either individual problems or across suites of problems. One branch of this field is hyper-heuristics, which operates on a space of low-level heuristics, searching for combinations of heuristics which exploit the strength and compensate for the weaknesses of individual known heuristics. The resulting algorithms are cheap to implement, require less human expertise, have robust performance within a problem class, and are portable across problem domains. These features compensate for some reduction in solution quality compared to tailor-made approaches, while still ensuring solutions of acceptable quality. However, most automated design approaches fail to incorporate or recognise a crucial human competence; human beings continuously learn from experience - by generalising observations and feedback, they are able to update their internal problem-solving models in order to continuously improve them, and adapt to changing circumstances. The failure of computational solvers to exploit previous knowledge both wastes useful knowledge and potentially hinders the discovery of good solutions. Furthermore, if the characteristics of instances of problems in the domain change over time, solvers may need to be completely re-tuned or in the worst case redesigned periodically.This proposal addresses these dual concerns raised above. We propose a novel lifelong-learning hyper-heuristic system which addresses current deficiencies inherent in current systems: it will exhibit short-term learning, producing fast and effective solutions to individual problems and at the same time, long-term learning processes will enable the system to autonomously adapt to new problem characteristics over time. It therefore exploits existing knowledge whilst simultaneously adapting to new information. Secondly, by working closely with two collaborators, a commercial routing software vendor and a forestry expert, our research will be directly informed by real-world problems, accounting for real constraints and performance criteria, thereby producing economic impact. Future advances in optimisation techniques will be facilitated by the development of a problem generator and a number of problem suites which reflect real-world priorities and constraints, derived from actual problem data provided through our collaborators and defined in conjunction with metrics which reflect not only economic drivers but also address environmental impact and the reduction of carbon emissions. This information database will be widely disseminated to provide an extensive platform for future research.