Dendritic cell algorithm
The dendritic cell algorithm is inspired by the danger theory of the mammalian immune system and, more specifically, the role and function of dendritic cells. The danger theory was proposed by Matzinger and suggests that the role of the acquired immune system is to respond to danger signals, rather than to distinguish self from non-self. The theory suggests that antigen-presenting cells (such as T helper cells) activate an alarm signal providing the necessary costimulation of antigen-specific cells to respond.
Dendritic cells are a type of cell in the innate immune system that respond to certain specific forms of danger signals. There are three main types of dendritic cells: immature which collect parts of the antigen and signals, semi-mature which are immature cells which decide internally that local signals represent security and present the antigen to T cells resulting in tolerance, and mature cells that internally decide that local signals represent danger and present the antigen to T cells resulting in a reactive response.
The information processing goal of the dendritic cell algorithm is to prepare a set of mature dendritic cells (prototypes) that provide context-specific information on how to classify normal and abnormal input patterns. This is achieved as a system of three asynchronous processes: 1) migration of sufficiently stimulated immature cells, 2) promotion of migrated cells to a semi-mature (safe) or mature (danger) state depending on their cumulative response , and 3) labeling the observed patterns safe or unsafe based on the composition of the subpopulation of cells that respond to each pattern.
The following algorithm provides a pseudocode for learning a pool of cells in the dendritic cell algorithm, in particular the deterministic dendritic cell algorithm. Mature migrated cells associate their collected input patterns with abnormalities, while semi-mature migrated cells associate their collected input patterns as normal. The resulting migrated cells can then be used to classify input patterns as normal or abnormal.
This can be done by sampling the cells and using a voting mechanism, or more elaborate methods such as a mature background antigen value (MCAV) which uses M/Ag (where M is the number of mature cells with antigen and Ag is the sum of antigen exposures by these mature cells), giving a probability that a pattern is an abnormality.
The dendritic cell algorithm is not specifically a algorithm classification, it can be considered as a data filtering method for use in anomaly detection problems. The canonical dendritic cell algorithm is designed to operate on a single discrete, categorical or ordinal input and two specific probabilistic signals indicating danger heuristic or entrance security.
Danger and safety signals are problem-specific signals of the risk of the input pattern being abnormal or normal, both usually in [0; 100]. Danger and safety signals do not have to be reciprocal, which means they can provide contradictory information. The system was designed for use in real-time anomaly detection problems, not just static problems. Each cell migration threshold is defined separately, usually in [5; 15].
