1,721,014 research outputs found
Emergent associative memory as a local organising principle for global adaptation in adaptive networks
No full textComplex adaptive systems composed of self-interested agents can in some circumstances self-organise into structures that enhance global adaptation or efficiency. However, the general conditions for such an outcome are poorly understood. In contrast, sufficient conditions for artificial neural networks to form structures that perform collective computational processes such as associative memory/recall, generalisation and optimisation, are well-understood. While such global functions within a single agent or organism may arise from mechanisms (e.g., Hebbian learning) that were selected for this purpose, agents in a multi-agent system have no obvious reason to produce such global behaviours when acting from individual interest. However, Hebbian learning is actually a very simple and fully-distributed habituation or positive feedback principle. Here we use an adaptive network model in which agents can modify their behaviours (states) but also their interactions with other agents (network topology). We show that when self-interested agents can modify how they are affected by other agents then, in adapting these inter- agent relationships to maximise their own utility, they will necessarily alter them in a manner homologous with Hebbian learning. When the agents adapt their behaviours relatively quickly, and their relationships with other agents relatively slowly, we find that the overall network dynamics are modified to find better adapted states more reliably. This separation in timescales causes the state dynamics to spend most of their time at attractors. Thus, the network develops an associative memory that amplifies a subset of its own attractor states. This self-organised modification to the network dynamics enhances its ability to resolve conflicts between agents. Moreover, we show that the system is not merely ‘recalling’ high quality states that have been previously visited, but ‘predicting’ their location by generalising over local attractor states that have already been visited. Thus, globally adaptive behaviours can emerge from self-organising adaptive networks that follow organisational principles familiar in connectionist models of organismic learning
How micro-evolution can guide macro-evolution: multi-scale search via evolved modular variation
No full textA divide-and-conquer approach to problem solving can in principle be far more efficient than tackling a problem as a monolithic whole. This type of approach is most appropriate when problems have the type of modular organisation known as near-decomposability, as implicit in many natural and engineered systems. Existing methods create higher scale composite units from non-random combinations of lower-scale units that reflect sub-problem optima. The use of composite units affords search at a higher scale that, when applied recursively, can ultimately lead to optimal top-level solutions. But for this approach to be efficient, we must decompose a problem in a manner that respects its intrinsic modular structure, information which is in general unavailable a priori. Thus, identifying and subsequently exploiting the structure recursively is vital in providing fully automatic problem decomposition.In this thesis, we define a family of algorithms that probabilistically adapt the scale of decomposition they use to reflect the structure in a problem. By doing so, they can provide optimisation that is provably superior to any single scale of search in nearly decomposable problems. Our proposed framework couples two adaptive processes: a rapid, fine-scale search that guides a slower adaptation of the decomposition. This results in a scaling up of the units used in the rapid search, now operating at a macro-scale. We find that separating the timescales for the fine-scale search and the adaptation of the decomposition is crucial for this kind of scalable optimisation. Using a simple and general class of problems that have no systematic structure, we demonstrate how our approach can nevertheless exploit the incidental structure present. Furthermore, we use idealised cases that have simple modular structure to demonstrate how our method scales as ?(N log N) (where N is the problem size), despite the fact that single-scale search methods scale as ? (2 ?N) – and support this distinction analytically.Although our approach is algorithmically superior to single-scale search, the underlying principles that it is constructed from are simple and can operate using only localised feedback. We discuss intriguing parallels between our approach and the significance of associative evolution for ecosystem adaptation. Our results suggest that macro-evolutionary processes might not be merely extended micro-evolution, but that the action of evolutionary processes upon several scales is fundamentally different from the conventional view of (micro-)evolution at a single scale
The Emperor’s New Clothes? An exploratory study into the purpose and effects of uniform with FE
No full textIntroduction
There is a long tradition of school uniform within the UK, yet the topic receives relatively little attention in UK-based research, particularly compared with other parts of the world. In my experiences within FE, I have encountered a prevailing narrative that uniform benefits students’ academic and vocational successes, yet I have not seen robust evidence to support this; whilst some studies explore such correlations, few – perhaps none – are situated within FE; yet this research setting has witnessed diminishing numbers of students adhering to uniform policy, against a backdrop of socioeconomic disruption caused at least in part by Covid-19. Accordingly, this research aims to explore purpose and effects of uniform in FE, with a view to informing future policy design within the setting.
Participants and Methods
The research was set in a large FE college in South-East England. Following an extensive literature review, which provides both justification and direction for the research, a particular faculty within the setting was selected, due to the breadth and variety of vocational courses delivered; all staff and students therein were invited to complete questionnaires, answering questions regarding motivation for uniform policy, understanding of policy, use of uniform, feelings towards uniform and perceived effects of uniform. 34 students and 7 staff responded. The questionnaires were a mixture of open ended and Likert-scale-type questions, producing both qualitative and quantitative data. Thematic analysis was applied to the qualitative data in order to produce semantic-based codes.
Results
The questionnaire results illustrated noticeable disparities between staff and students, regarding their respective understanding of uniform, the reasons for such policies and the desire for them. Neoliberalist values were apparent in the staff’s responses, whilst Foucaultian principles were evident not only within the literature review but within college narratives too.
Discussion
I explore emerging tensions within the setting and suggest possible reprecussions for not addressing these issues promptly. I acknowledge particular implications of uniform relating to cost and inclusion, noting injustices felt by some students, and suggest that policy makers would benefit from including both students and employers in their policy-design processes; such collaboration may lead to greater understanding by all stakeholders, of necessary vocational development and its relationship with uniform. Additionally, I note some limitations of the research and make recommendations for future research
Multi-scale search, modular variation, and adaptive neighbourhoods
No full textThis paper investigates a framework for multi-scale search, which makes use of automatically defined multi-variable groupings to search over multiple scales of organisation. We show how this method redefines the variational neighbourhood visible to the search process to efficiently find high quality optima. We use a transparent two-scale modular problem to provide a clear explanation of when and how this approach works, supporting this with analysis of expected runtimes. This problem has an obvious and familiar modular structure but, unlike previous test problems used for existing model-building methods, the size of the modules are not constant but scale as a function of problem size. We show that this problem class requires multi-scale search to be solved in polynomial time
Symbiosis Enables the Evolution of Rare Complexes in Structured Environments
No full textWe present a model that considers evolvable symbiotic associations between species, such that one species can have an influence over the likelihood of other species being present in its environment. We show that this process of 'symbiotic evolution' leads to rare and adaptively significant complexes that are unavailable via non-associative evolution
Transformations in the Scale of Behaviour and the Global Optimisation of Constraints in Adaptive Networks
No full textThe natural energy minimisation behaviour of a dynamical system can be interpreted as a simple optimisation process, finding a locally optimal resolution of problem constraints. In human problem solving, high-dimensional problems are often made much easier by inferring a low-dimensional model of the system in which search is more effective. But this is an approach that seems to require top-down domain knowledge; not one amenable to the spontaneous energy minimisation behaviour of a natural dynamical system. However, in this paper we investigate the ability of distributed dynamical systems to improve their constraint resolution ability over time by self-organisation. We use a ‘self-modelling’ Hopfield network with a novel type of associative connection to illustrate how slowly changing relationships between system components can result in a transformation into a new system which is a low-dimensional caricature of the original system. The energy minimisation behaviour of this new system is significantly more effective at globally resolving the original system constraints. This model uses only very simple, and fully-distributed positive feedback mechanisms that are relevant to other ‘active linking’ and adaptive networks. We discuss how this neural network model helps us to understand transformations and emergent collective behaviour in various non-neural adaptive networks such as social, genetic and ecological networks
Adaptive units of selection can evolve complexes that are provably unevolvable under fixed units of selection (abstract)
No full textSymbiosis, the collaboration of multiple organisms from different species, is widespread amongst prokaryotes. As symbiotic associations become more absolute, an inseparable union can result (symbiogenesis). This union would typically be considered a new unit of selection once a set of symbionts became reproductively inseparable. However we consider a macroevolutionary model within which new units of selection are formed gradually from maturing symbiotic relationships, without the need for symbiogenic events to be complete. We find that these adaptive units of selection can evolve complexes that are provably unevolvable under fixed units of selection. We perform experiments in several structured rugged landscapes in which there only few global optima amongst many local optima. Importantly, the number of initial conditions that must be sampled to find one of these global optima by individual adaptation is exponentially large (i.e., their basins of attraction are small). Selection on individual species only has the ability to evolve to locally optimal configurations. Selecting on groups at the level of entire ecosystems requires an exponentially large number of ecosystems to be created to find optima of globally maximal fitness. Since this level of selection cannot utilise gradient information, global optima are only guaranteed to be found by enumerating all possible ecosystem configurations. In our model the ecosystem is sub-divided into several semi-independent demes, each of which evolves to local attractors via individual adaptation. Symbiotic associations between the entities that are present in the local attractor discovered by each deme are modified to reflect their frequency of co-occurrence in the attractors across all of the demes. These two phases are repeated, resulting in coalitions of increasing size, stability, and co-dependence. We compare two variants for how the symbiotic associations are formed: associations between entities that are ‘all-or-nothing’, and probabilistic associations. Selection starts at the level of the individual, when local optima are discovered. However, the unit of selection adapts as composites gradually develop. Some composites are favoured at the expense of other composites, leading to the discovery of higher fitness optima. These composites consist of co-adapted groups that are locally optimal under individual selection, and not random configurations of large sets of entities. Thus, although the groups can contain many entities, the number of competing complexes is small: the configuration space is reduced as symbiotic relationships are strengthened. We demonstrate the adaptive significance of these processes by applying them to provably difficult optimisation problems. In general, optimisation problems frequently have some structure that can be exploited, but only with an appropriate mechanism that can recognise that structure. When locally optimal configurations contain some information that is congruent with globally optimal solutions, such as when local optima are created from the frustration of a large number of low-order fitness interactions, the process described above can provide automatic problem decomposition. When composite entities evolve through the intensification of symbiotic coalitions, the units of selection adapt causing the configuration space dimensionality to be collapsed. We show that this can result in complexes that are unevolvable when the units of selection are static
Atomic computing - a different perspective on massively parallel problems
No full textAs the size of parallel computing systems inexorably increases, the proportion of resource consumption (design effort, operating power, communication and calculation latency) absorbed by ‘non-computing’ tasks (communication and housekeeping) increases disproportionally. The SpiNNaker (Spiking neural net architecture) engine [1,2] sidesteps many of these issues with a novel architectural model: it is an isotropic ‘mesh’ of (ARM9) cores, connected via a hardware communication network. The topology allows uniform scalability up to a hard limit of just over a million cores, and the communications network – hardware handling packets of 72 bits – achieves a bisection bandwidth of 5 billion packets/s. The state of the machine is maintained in over 8TB of 32-bit memory, physically distributed throughout the system. There is no central processing ‘overseer’ or synchronised clock. This paper discusses opportunities and challenges in applying the SpiNNaker architecture, within neural simulation and beyond
- …
