Alwyn V. Husselmann

Geometric Firefly Algorithms on Graphical Processing Units

Geometric unification of Evolutionary Algorithms (EAs) has resulted in an expanding set of algorithms which are search space invariant. This is important since search spaces are not always parametric. Of particular interest are combinatorial spaces such as those of programs that are searchable by parametric optimisers, providing they have been specially adapted in this way. This typically involves redefining concepts of distance, crossover and mutation operators. We present an informally modified Geometric Firefly Algorithm for searching expression tree space, and accelerate the computation using Graphical Processing Units. We also evaluate algorithm efficiency against a geometric version of the Genetic Programming algorithm with tournament selection. We present some rendering techniques for visualising the program problem space and therefore to aid in characterising algorithm behaviour.

INTELLIGENT INDIVIDUAL AGENT-BASED SIMULATION OF PHOTOBIOREACTORS AND GROWTH CONTROL

Photobioreactors offer an efficient and controllable environment from which many biomaterial-based products can be manufactured naturally. For example, the cultivation of the algae H. pluvialis, results in a carotenoid known as Astaxanthin, which is a highly valuable pigment and powerful antioxidant. The demand for naturally-produced Astaxanthin is much higher than synthetic, and the process by which this algae is cultivated is therefore important. We develop a discrete agent-based model that can be used to test the control processes for growing and harvesting such a bio product. We experiment with a variant of the Photosynthetic Factory model to improve cell division and hence growth kinetics. We employ parallel computing techniques to implement simulations of large numbers of control agents on graphical processing units. We discuss the implications of stochasticity, and propose a metric with which to balance cultivation time against fill rates of a stochastic data-parallel, bubble-column photobioreactor simulation.

Simulating growth kinetics in a data-parallel 3d lattice photobioreactor

Though there have been many attempts to address growth kinetics in algal photobioreactors, surprisingly little have attempted an agent-based modelling (ABM) approach. ABM has been heralded as a method of practical scientific inquiry into systems of a complex nature and has been applied liberally in a range of disciplines including ecology, physics, social science, and microbiology with special emphasis on pathogenic bacterial growth. We bring together agent-based simulation with the Photosynthetic Factory (PSF)model, as well as certain key bioreactor characteristics in a visual 3D, parallel computing fashion. Despite being at small scale, the simulation gives excellent visual cues on the dynamics of such a reactor, andwe further investigate the model in a variety of ways. Our parallel implementation on graphical processing units of the simulation provides key advantages, which we also briefly discuss. We also provide some performance data, along with particular effort in visualisation, using volumetric and isosurface rendering.

RANDOM FLIGHTS FOR PARTICLE SWARM OPTIMISERS

Parametric Optimisation is an important problem that can be tackled with a range of bio-inspired problem space search algorithms. We show how a simplified Particle Swarm Optimiser (PSO) can efficiently exploit advanced space exploration with L´ evy flights, Rayleigh flights and Cauchy flights, and we discuss hybrid variations of these. We present implementations of these methods and compare algorithmic convergence on several multi-modal and unimodal test functions. Random flights considerably enhance the efficient simplified PSO and the L´ evy flight gives good balance between local space exploration and local minima avoidance. We discuss computational tradeoffs involved in generating such flights. In summary, these modifications show varying success between themselves for problem solving, but outperforms the uniform random exploration technique in most cases.

Particle Swarm-based Meta-Optimising on Graphical Processing Units

Optimisation (global minimisation or maximisation) of complex, unknown and non-differentiable functions is a dif- ficult problem. One solution for this class of problem is the use of meta-heuristic optimisers. This involves the system- atic movement of n-vector solutions through n-dimensional parameter space, where each dimension corresponds to a parameter in the function to be optimised. These meth- ods make very little assumptions about the problem. The most advantageous of these is that gradients are not neces- sary. Population-based methods such as the Particle Swarm Optimiser (PSO) are very effective at solving problems in this domain, as they employ spatial exploration and local solution exploitation in tandem with a stochastic compo- nent. Parallel PSOs on Graphical Processing Units (GPUs) allow for much greater system sizes, and a dramatic reduc- tion in compute time. Meta-optimisation presents a further super-optimiser which is used to find appropriate algorith- mic parameters for the PSO, however, this practice is often overlooked due to its immense computational expense. We present and discuss a PSO with an overlaid super-optimiser also based on the PSO itself.

Model structure optimisation in lattice-oriented agent-based models

Agent-based models (ABMs) have shown themselves to be very suitable for examining complex phenomena. Their rise in popularity and multi-disciplinary interest has spurred many researchers to seek more efficient algorithms and methodologies, such as parameter optimisation, interaction redundancy reduction, model induction and meta-modelling. Some of these efforts are directed towards reducing the difficulty associated with them, which has limited the wider adoption of ABM. Several Domain-specific languages (DSLs) have been developed, many of which ease the difficulty faced by programmers from disciplines unrelated to computer science. Aside from the need of knowledge in programming, some researchers have also expressed a need to model larger systems. These can easily become too computationally expensive, and more difficult to build and manage. In this work, we propose a simple extensible DSL which performs behaviour generation using domain-specific knowledge, whilst avoiding run-time interpretation and abstraction penalties. This is done using a simple evolutionary optimiser, which operates directly on program syntax trees, computing recombinations. We allow the user to explore uncertainty and discover models in a guided manner; as opposed to tedious trial-and-error. Our DSL makes use of the Multi-Stage Programming paradigm within Terra, in order to just-in-time compile generated programs at run-time, as well as allowing the optimiser to modify program syntax trees. The goal of the optimiser is to reduce uncertain syntax constructs to concrete ones while exploring the search space of a given objective function. This results in a guided modelling language with high performance and extensibility. We discuss the implementation of this language, with its compiler architecture, as well as the incorporation of the evolutionary optimiser, and we demonstrate the use of this approach in solving the Santa Fe Ant Trail problem.

K-expressions visualisation

Visualisation is important for gaining a qualitative understanding of how algorithms operate [1, 2]. Visualization techniques have been used to shed light on 3D voxel sets [3], vector fields [4], as well as spatial data structures [5], lattice gases [6], and several evolutionary algorithms.

MULTI-STAGE, HIGH PERFORMANCE, SELF-OPTIMISING DOMAIN-SPECIFIC LANGUAGE FOR SPATIAL AGENT-BASED MODELS

Optimisation in the context of Agent-based Modelling has been thoroughly researched and reported in the literature. In particular, model parameter tuning has been done using a variety of parametric optimisers, and we are now entering a phase where agent behaviour itself is learned, not specified. The latter is proving to be problematic for a number of reasons. Algorithms earmarked for this purpose such as Genetic Programming and decision tree induction present their own problems. Defining the search space in terms of building blocks for these algorithms is surprisingly difficult. We propose a different methodology for accomplishing machine learning in the context of model induction. Instead of forcing the modeller to provide fine grained and concise model building blocks, we provide a language where small portions of uncertain dynamics can be expressed concisely using domain specific knowledge. This has the potential to greatly increase the efficiency of building simulations for models, and reduce time spent on verification. Our language is built using recent concepts of multi-stage programming (MSP), providing run-time compiling and execution of code. This allows us to avoid the

On the Robustness of Fuzzy-Genetic Colour Contrast Fusion with Variable Colour Depth

Present in this paper is a hybrid Fuzzy-Genetic colour classification system that works under spatially varying illumination intensities, even for moving source illuminants, in real-time. At the heart of the system is an algorithm called Fuzzy Colour Contrast Fusion (FCCF) that compensates for all confounding effects in the environment. We extended FCCF to become self-calibrating using a Heuristic-Assisted Genetic Algorithm (HAGA), and enhanced it further using a technique called Variable Colour Depth (VCD). We devised an improved fitness function for finding the best colour contrast rules and compared it with the rule scoring system used previously by FCCF. Moreover, we tested the integrated algorithms on the FIRA robot soccer platform, but with much more challenging lighting conditions. Our experiments include real-time colour object recognition under extreme illumination conditions, such as, multiple source illuminants, arbitrarily moving source illuminant and colour classification under environments not seen during training. Our results attest to the robustness of the proposed hybrid system, with colour object recognition accuracy ranging from 84% to 100%, measured as robot recognition per frame.