Immune Systems - an evolutionary metaphor

Immune Systems - an evolutionary metaphor Steve Cayzer BICAS research group Hewlett-Packard Laboratories, Bristol February 2003 AIS – 2 minute overview • The human body is constantly under attack from antigens (foreign proteins) The immune system creates antibodies which match and destroy these antigens … without destroying the host (self proteins) • • • Each antibody matches a range of proteins: as a population, antibodies (learn to) cover non-self space. Adaptive, self organising system: good BICAS paradigm • 30/04/2008 Immune Systems - an evolutionary metaphor page 2 of 25 Antibodies map non-self space Non-Self Self X X X X X X X X X Antigen Antibody (with recognition radius) (matched by antibody) 30/04/2008 Immune Systems - an evolutionary metaphor page 3 of 25 Features and Applications Features     Applications     Learning & Adaptation Immunological Memory Self/Non self classification Security Pattern Recognition Heuristic Optimisation Modelling Studies Associative Recall/SelfOrganizing Localization/Circulation (Island model)   Agents    Clustering Concept Learning Recommender Systems  Autonomous, decentralized 30/04/2008 Immune Systems - an evolutionary metaphor page 4 of 25 AIS – Basic Algorithm Initialise antibody population WHILE (not finished) Present antigens Calculate immune response (matching) Propagate effect to (idiotypic) network Lifecycle events - creation - destruction - screening END WHILE 30/04/2008 Immune Systems - an evolutionary metaphor page 5 of 25 Philosophical Divide The IS can be thought of as a special case of: Genetic algorithm     Neural network     Creation (gene libraries) Emphasis on mutation Matching ~ fitness (?) Variable population size Pattern classification Unsupervised learning Topographic mapping Variable network topology Considerations   Considerations   Role of antigen Preservation of diversity AIS as optimiser Interpreting response Training regime AIS as classifier 30/04/2008 Immune Systems - an evolutionary metaphor page 6 of 25 Evolutionary AIS metaphors • Optimisation • Constraint Handling • Scheduling • Partial solutions • Niching and Diversity • AIS and GP • Co-evolutionary approaches 30/04/2008 Immune Systems - an evolutionary metaphor page 7 of 25 AIS for Optimisation  Antibodies as: entire solutions „building blocks‟ as: objective functions constraints fit/feasible solutions solutions to subproblems weight combinations spanning Pareto optimal front  Antigens  AIS usually hybridised with GA: Antibody selection Gene library creation  Emergent fitness sharing (generalist/specialist) 30/04/2008 Immune Systems - an evolutionary metaphor page 8 of 25 Example: Hajela & Yoo 2001 Designs Feasible Infeasible Antigen Antibody Crossover Small s all Generalist AIS best Mutation GA (unconstrained objective function) 30/04/2008 Immune Systems - an evolutionary metaphor page 9 of 25 AIS for Optimisation Evaluation to TSP (of course), job shop scheduling, time series prediction, truss design, capacitor placement, time dependent optimization…   Applied Some good results on test problems BUT…  Often little „added value‟ to GA metaphor somewhat strained  AIS   Difficult to find fair comparisons Best viewed as a collection of hybridising techniques Immune Systems - an evolutionary metaphor page 10 of 25 30/04/2008 AIS and document classification • AIS appears to perform some form of classification (self/non-self) We can apply this to web-based document filtering (interesting/not interesting) The idea is to build an AIS with antibodies that will „recognise‟ interesting documents • • • Uses coevolutionary learning Outperforms traditional paradigms Twycross & Cayzer 2003 An immune-based approach to document classification (IIPWM 2003; available at http://www.hpl.hp.com/techreports/2002/HPL-2002-292.html) • 30/04/2008 Immune Systems - an evolutionary metaphor page 11 of 25 AIS as a coevolutionary concept learner Antibody Species … AB … … … “Best” Antibody BC CDE AE Antibody Serum (= concept) (AB)  (BC)  (CDE)  (AE) …where A, B etc are document features (eg keywords) 30/04/2008 Immune Systems - an evolutionary metaphor page 12 of 25 AIS Classification Results 30/04/2008 Immune Systems - an evolutionary metaphor page 13 of 25 Future Prospects • • • • • Optimization: has taken a back seat. Data Mining: classification, clustering AIS for security: research ongoing Idiotypic, self organizing flavour (community?) Danger Theory: EPSRC „adventure fund‟ proposal Immune Systems - an evolutionary metaphor page 14 of 25 30/04/2008 Pause for questions This page intentionally left blank 30/04/2008 Immune Systems - an evolutionary metaphor page 15 of 25 Other cool stuff 1. 2. 3. 4. Other immune system elements Refinements to the basic algorithm The Idiotypic Effect The Danger Theory (goto_beer) 30/04/2008 Immune Systems - an evolutionary metaphor page 16 of 25 A slightly less simplified immune system Innate vs Acquire d Humor al (back) Cell Mediated vs T Cell (CD-4, Helper) Binds to MHC-antigen complex Secretes cytokines to help… B Cell Secretes Antibody T Cell (CD-8, Killer) Kills cell (viruses) which binds to antigen and recruits phagocytes (innate) 30/04/2008 Immune Systems - an evolutionary metaphor page 17 of 25 AIS – Basic Algorithm Initialise antibody population WHILE (not finished) Present antigens Calculate immune response (matching) Propagate effect to (idiotypic) network Lifecycle events - creation - destruction - screening END WHILE 30/04/2008 Immune Systems - an evolutionary metaphor page 18 of 25 AIS – Refined algorithm Basic Matching Algorithm   Population of B Cells (antibodies) Clonal expansion and hypermutation Extensions     Lifecycle events, screening (positive/negative selection) Other IS elements (T Cells, cytokines) Network interactions (idiotypic effects) Other – localization, self adaptation, population control Choices   Genotype/Phenotype (Representation & Shape Space) Matching (Hamming, Euclidean, r-contiguous, other) Immune Systems - an evolutionary metaphor (back) page 19 of 25 30/04/2008 The Idiotypic Effect: Antibody-antibody interactions Internal Image of Antigen Anti-Idiotypic Set P2 - Jerne’s Big Idea (1974) Idiotype: specificity of antibody (epitopes to which it will bind) Idiotope: An idiotypic epitope Evidence: Antibodies produced against antibodies of same species (cf individual) I2 P3 + I3 I1 Idiotypic Set P1 Antigen 30/04/2008 Immune Systems - an evolutionary metaphor page 20 of 25 The Idiotypic Effect – Why do we care? • • • Biological importance - ??? Immunological models – Varela, Castellani Pattern recognition – Timmis & Hunt • Non-stationary environments (idiotypic memory) – Gaspar & Collard Multimodal Optimisation – de Castro Recommendation communities – Cayzer & Aickelin Immune Systems - an evolutionary metaphor page 21 of 25 • • 30/04/2008 Modelling the Idiotypic Effect dxi dt  antibodies   I am   antigens   death   c   recognised    recognised    recognised    rate                 N n N   c  m ji xi x j  k1  mij xi x j   m ji xi y j   k 2 xi j 1 j 1  j 1  • • • • For N antibodies, n antigens. xi is the concentration of antibody i yi is the concentration of antigen I c, k1 and k2 are scaling constants (back) page 22 of 25 • mij is a matching function 30/04/2008 Immune Systems - an evolutionary metaphor Problems with the self-nonself worldview • How do we produce antibodies that react against antigens and yet avoid self? One way is “Generate and Test”: negatively screen antibodies which react to self at production time But this is expensive! • • • • • It‟s difficult to screen against ALL self. Self also changes over time And it is not necessary to screen against all non-self – only dangerous non-self Aickelin & Cayzer 2002 The Danger Theory and Its Application to Artificial Immune Systems Proc. International Conference on AIS (ICARIS 2002) 30/04/2008 Immune Systems - an evolutionary metaphor page 23 of 25 The Danger Theory • In the danger model, the idea is to recognise „danger‟ rather than non self. The screening is accomplished post production through an external „danger‟ signal. • • Thus the production of autoreactive antibodies (which react to self) is allowed. If an (eg autoreactive) antibody matches a stimulus in the absence of danger, it is removed. Thus harmless antigens are tolerated, and changing self accommodated. Matzinger 2002 The Danger Immune Systems -renewed metaphor of self Science 296: 301-304 Model: A an evolutionary sense page 24 of 25 • • 30/04/2008 Potential Implications of the Danger Theory • In computer security, we need to discriminate between safe and dangerous activity. In the “Generate and Test” paradigm, detectors (eg for network traffic patterns) are screened against „normal‟ activity. „Danger‟ signals such as memory usage, SIGABRT signals etc. could be useful evidence to help a security AIS refine its detectors. It could also be useful for data mining, where the „danger‟ signal is a proxy measure of interest EPSRC Adventure Fund proposal in progress… • • • • (back) 30/04/2008 Immune Systems - an evolutionary metaphor page 25 of 25