Dendritic cell algorithm
The dendritic cell algorithm is inspired by the theory of danger 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 distinguishing self from non-self. The theory suggests that antigen presenting cells (such as helper T cells) activate an alarm signal providing the necessarily co-stimulation of antigen-specific cells to respond.
Dendritic cells are a type of cells in the innate immune system that respond to 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 internally decide that local signals represent safety and present the antigen to the resulting T cells. tolerance, and mature cells that internally decide that local signals are dangerous and present the antigen to T cells resulting in a reactive response.
The information processing objective 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 safe or unsafe patterns 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 sophisticated methods such as a mature context antigen (MCAV) value 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), which gives a probability that a pattern is an abnormality.
The dendritic cell algorithm is not specifically a algorithm of 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 signals specific to the problem of the risk that the input pattern is an anomaly or is normal, both typically in [0; 100]. Danger and safety signals do not have to be reciprocal, which means they can provide conflicting information. The system was designed to be used in real-time anomaly detection problems, not just static problems. Each cell migration threshold is defined separately, usually in [5; 15].