Deep Optimisation: Transitioning the Scale of Evolutionary Search by Inducing and Searching in Deep Representations

Abstract

We investigate the optimisation capabilities of an algorithm inspired by the Evolutionary Transitions in Individuality. In these transitions, the natural evolutionary process is repeatedly rescaled through successive levels of biological organisation. Each transition creates new higher-level evolutionary units that combine multiple units from the level below. We call the algorithm Deep Optimisation (DO) to recognise both its use of deep learning methods and the multi-level rescaling of biological evolutionary processes. The evolutionary model used in DO is a simple hill-climber, but, as higher-level representations are learned, the hill-climbing process is repeatedly rescaled to operate in successively higher-level representations. The transition process is based on a deep learning neural network (NN), specifically a deep auto-encoder. Our experiments with DO start with a study using the NP-hard problem, multiple knapsack (MKP). Comparing with state-of-the-art model-building optimisation algorithms (MBOAs), we show that DO finds better solutions to MKP instances and does so without using a problem-specific repair operator. A second, much more in-depth investigation uses a class of configurable problems to understand more precisely the distinct problem characteristics that DO can solve that other MBOAs cannot. Specifically, we observe a polynomial vs exponential scaling distinction where DO is the only algorithm to show polynomial scaling for all problems. We also demonstrate that some problem characteristics need a deep network in DO. In sum, our findings suggest that the use of deep learning principles have significant untapped potential in combinatorial optimisation. Moreover, we argue that natural evolution could be implementing something like DO, and the evolutionary transitions in individuality are the observable result.

The collective intelligence of evolution and development

“…The ability of distributed learning to improve problem-solving ability in this way is now well-developed (Kounios et al, 2016; Mills, 2010; Mills et al, 2014; Watson et al, 2011a; Watson et al, 2011b, 2011c; Watson et al, 2016). In some conditions, a learning neural network can enable a sort of ‘chunking’, rescaling the search process to a higher level of organisation (Caldwell et al, 2018; Mills, 2010; Watson et al, in review; Watson et al, 2011c; Watson et al, 2016). Elsewhere, we hypothesise that this rescaling of the problem-solving search process is intrinsic to transitions in individuality (Watson et al, 2016), suggesting that ETIs constitute a form of deep model induction (Czegel et al, 2019; Vanchurin et al, 2021; Watson et al, 2022).…”

“…They fall short, however, of demonstrating the spontaneous evolution of a new level of individuality. In algorithmic terms, such models cannot do the ‘chunking’ of the search space or rescaling of the search process that is facilitated by the induction of deep models (Caldwell et al, 2018; Mills, 2010; Mills et al, 2014; Watson et al, 2011b; Watson et al, 2016; Watson et al, 2009). We hypothesise that this is because they are single-level networks of symmetric interactions; our roadmap supports the idea that the evolutionary transitions in individuality correspond to deep interaction structures (Czegel et al, 2019; Watson et al, 2022) or perhaps other mechanisms of multi-scale dynamics (Watson, accepted; Watson et al, in review).…”

Evolvability and organismal architecture

“…No one to my knowledge has ever explicitly made the connection between these two ideas until now. However, it has to be said that autoencoders are being used in evolutionary computation already and that Richard Watson as well has been recently working with this representation [127].…”