Chris McEwan

Representation in the (Artificial) Immune System

Much of contemporary research in Artificial Immune Systems (AIS) has partitioned into either algorithmic machine learning and optimisation, or, modelling biologically plausible dynamical systems, with little overlap between. We propose that this dichotomy is somewhat to blame for the lack of significant advancement of the field in either direction and demonstrate how a simplistic interpretation of Perelson’s shape-space formalism may have largely contributed to this dichotomy. In this paper, we motivate and derive an alternative representational abstraction. To do so we consider the validity of shape-space from both the biological and machine learning perspectives. We then take steps towards formally integrating these perspectives into a coherent computational model of notions such as life-long learning, degeneracy, constructive representations and contextual recognition—rhetoric that has long inspired work in AIS, while remaining largely devoid of operational definition.

On AIRS and Clonal Selection for Machine Learning

AIRS is an immune-inspired supervised learning algorithm that has been shown to perform competitively on some common datasets. Previous analysis of the algorithm consists almost exclusively of empirical benchmarks and the reason for its success remains somewhat speculative. In this paper, we decouple the statistical and immunological aspects of AIRS and consider their merits individually. This perspective allows us to clarifying why AIRS performs as it does and identify deficiencies that leave AIRS lacking. A comparison with Radial Basis Functions suggests that each may have something to offer the other.

Advances in artificial immune systems

The field of Artificial Immune Systems (AIS) derives inspiration from processes and mechanisms apparent in the biological immune system. After early applications of this paradigm to problems in anomaly detection and classification, the field rapidly expanded, leading to a number of successful optimization algorithms, and more recently, applications in fields as diverse as swarm robotics and wireless sensor networks. In contrast to the situation in most other areas of biologically-inspired computing, practitioners of AIS maintain close links with the field that inspires their work. The field of immunology itself is continually moving, with new discoveries constantly refining the knowledge and understanding we have of the natural immune system. New immunological knowledge is translated to AIS in the form of new algorithms and architectures, but in addition, research driven by computational modelling also informs immunology. This two-way flow of knowledge between engineering and immunology distinguishes the field from other biological paradigms in computing, where the underlying biological phenomena are perhaps more clearly understood. In parallel to this, a body of literature has now emerged which provides a theoretical underpinning for the field, employing many of the same techniques that are used to analyze other randomized search algorithms. This facilitates a more rigorous comparison between AIS algorithms and other search algorithms, and further enables crossover of ideas between AIS and other paradigms. The papers in this special issue—selected from papers submitted following the 9th International Conference on Artificial Immune Systems (ICARIS) held in Edinburgh in July 2010—capture the essence of the current state of the art in the field, illustrating the aforementioned trends. In their paper Collective Classification of Textual Documents by Guided Self-Organization in T-Cell Cross-regulation Dynamics, Abi-Haidar and Rocha analyse a computational extension of the model of T-Cell cross-regulation initially proposed by Carneiro et al. This work provides several interesting contributions to the AIS literature. On the one hand, it elaborates the original, minimalist analytical model to incorporate heterogeneous surface presentation and T-Cell specificities, which in turn, provides insight into how the hypothesised mechanism might unfold in a more realistic setting. On the other hand, this demonstration is provided in the context of document classification, where amino-acid sequences are replaced with words and tolerance and immunity replaced with skewed, two-class classification. The subsequent empirical demonstration of the model’s performance in comparison to established statistical algorithms provides a compelling insight into the robustness of the biological model that would be very difficult to assert using traditional means, due to the lack of sufficient experimental data. In Surrogate-assisted Clonal Selection Algorithms for Expensive Optimization Problems, Bernardino et al. provide a thorough empirical assessment of using surrogate functions in the CLONALG optimization algorithm. Stochastic search algorithms, of which CLONALG is an immune-inspired variant, often require a large amount of function evaluations. Surrogate functions provide a lightweight proxy objective function that can be used in place of the true underlying objective function, which can be exploited in several ways. Bernardino et al. show how surrogate functions can be used to improve solutions given the constraint of a fixed number of true objective evaluations. Such a constraint is typical of inverse problems in scientific computing, where the objective function may be an expensive simulation. Their paper demonstrates the applicability of these techniques to scale immune-inspired optimisation algorithms to this important class of problems, by comparison with a default CLONALG-based approach. They also provide some empirical insight into the best method of approximating the underlying objective function by comparing nearest-neighbour and locally linear variants. Finally, in On the Role of Age Diversity for Effective Aging Operators, Jansen and Zarges perform a theoretical study of the aging mechanisms in randomized search heuristics. Aging is a procedure employed in both evolutionary computation and AIS algorithms, whereby each search point is assigned an age which increases with each generation, and when a search point becomes older than some pre-defined maximal age, it is deleted from the population. New search points, or offspring, are then introduced, with the overall effect being to increase the diversity of the population, but there are various different strategies for assigning an age to the offspring. This paper considers different aging operators and provides asymptotic bounds on their performance when applied to optimization of a particular example objective function. These theoretical bounds are also compared with experimental results for small problem sizes, revealing a gap between theory and practice. The overall conclusion is that not only diversity of search points, but also diversity with respect to age plays an important role in determining the performance of randomized search algorithms that employ aging strategies.

On clonal selection

Clonal selection has been a dominant theme in many immune-inspired algorithms applied to machine learning and optimisation. We examine existing clonal selection algorithms for learning from a theoretical and empirical perspective and assert that the widely accepted computational interpretation of clonal selection is compromised both algorithmically and biologically. We suggest a more capable abstraction of the clonal selection principle grounded in probabilistic estimation and approximation and demonstrate how it addresses some of the shortcomings in existing algorithms. We further show that by recasting black-box optimisation as a learning problem, the same abstraction may be re-employed; thereby taking steps toward unifying the clonal selection principle and distinguishing it from natural selection.

Exploiting Collaborations in the Immune System: The Future of Artificial Immune Systems

Despite a steady increase in the application of algorithms inspired by the natural immune system to a variety of domains over the previous decade, we argue that the field of Artificial Immune Systems has yet to achieve its full potential. We suggest that two factors contribute to this; firstly, that the metaphor has been applied to insufficiently complex domains, and secondly, that isolated mechanisms that occur in the immune system have been used naively and out of context. We outline the properties of domains which may benefit from an immune approach and then describe a number of immune mechanisms and perspectives that are ripe for exploration from a computational perspective. In each of these mechanisms collaboration plays a key role. The concepts are illustrated using two exemplars of practical applications of selected mechanisms from the domains of machine learning and wireless sensor networks. The article suggests that exploiting the collaborations that occur between actors and signals in the immune system will lead to a new generation of engineered systems that are fit for purpose in the same way as their biological counterparts.

Clonal selection from first principles.

Clonal selection is the keystone of mainstream immunology and computational systems based on immunological principles. For the latter, clonal selection is often interpreted as an asexual variant of natural selection, and thus, tend to be variations on evolutionary strategies. Retro-fitting immunological sophistication and theoretical rigour onto such systems has proved to be unwieldy. In this paper we assert the primacy of competitive exclusion over selection and mutation; providing theoretical analysis and empirical results that support our position.