Mundane Rationality as a basis for modelling and understanding behaviour within specific contexts

Mundane Rationalityas a basis for modelling and understanding behaviour within specific contextsBruce EdmondsCentre for Policy ModellingManchester Metropolitan University

Social Intelligence Hypothesis (SIH)Kummer, H., Daston, L., Gigerenzer, G. and Silk, J. (1997)The crucial evolutionary advantages that human intelligence gives are due to the social abilities and structures it facilitatesThis explains the prevalence of specific abilities such as: imitation, language, social norms, lying, alliances, gossip, politics etc.Social intelligence is not a result of general intelligence applied to social organisation, but the essential core of human intelligencein fact our “general” intelligence could be merely a side-effect of social intelligenceMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 2

Implications of SIHThat different complex “cultures” of knowledge are significantAn important part of those cultures is how to socially organise, behave, coordinate etc.One should expect different sets of social knowledge for different groups of peopleThat these might not only be different in terms of content but imply very different ways of coordinating, negotiating, cooperating etc.That these will relate as a complete “package” to a significant extent, that has developed over time and passed down to new membersMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 3

An Evolutionary PerspectiveSocial intelligence implies that:Groups of humans can develop their own, very different, (sub)cultures of technologies, norms etc. (Boyd and Richerson 1985)These allow the group with their culture to inhabit a variety of ecological niches (e.g. the Kalahari, Polynesia) (Reader 1980)Thus humans, as a species composed of separate groups with different cultures and survival strategies, are able to survive catastrophes that effect different niches in different ways (specialisation)Mundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 4

Social Embedding (SE)Granovetter (1985) AJS 91 (3): 481-510Contrasts with the under- and over-socialised models of behaviourThat the particular patterns of social interactions between individuals matterIn other words, only looking at individual behaviour or aggregate behaviour misses crucial aspectsThat the causes of behaviour might be spread throughout a society – “causal spread”Shown clearly in some simulation modelsMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 5

Illustration of Causal ComplexityLines indicate causal link in behaviour over time, each box an agent’s talk or action decision (Edmonds 1999)Mundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 6

Implications of Social EmbeddingIn many circumstances agents can learn to exploit the particular computation and knowledge in their society, rather than do it themselves (invest in what Warren Buffet invests in)This knowledge is often not explicit but is something learned – this takes timeThis is particularly true of social knowledge – studying guides as to living in a culture are not the same as living there for a timeTrying to make social knowledge explicit, rather than adapt to it may be infeasible due to the complexity of the social embeddingOur personal networks of friends and colleagues becomes our extended social bodyMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 7

Context-dependency of Cognition“The” context is the situation of an event, but this is indefinitely extensiveThe brain somehow categorises and recognises different kinds of situation and preferentially gives access to knowledge on this basis, it is context-dependentMany aspects of human cognition seem to be context-dependent, including: memory, visual perception, choice making, reasoning, emotion, and languageMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 8

The Context HeuristicThe kind of situation is recognised in a rich, fuzzy, complex and unconscious mannerKnowledge, habits, expectations etc. are learnt for that kind of situation and are retrieved for itReasoning, learning, interaction happens with respect to the recognised kind of situationThese learnt kinds of situation are socially co-developed in time becoming entrenched in society and passed down the generationsFor example: lectures, interviews, partiesNot a general heuristic, but one particularly suited to the complexity of a socially constructed environmentMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 9

The general idea for context-dependent intelligencefuzzy but rich learning & recognition of the contextActionPerceptionprecise reasoning & belief update within contextMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 10

M1M2M1Abstract to a contextClusters of Domain and Content make a ContextMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 11

Implications of Context-DependencyBehaviour of observed actors might need to change sharply across different social contextsThe relevant behaviour, norms, kinds of interaction etc. might also need to changeSocial contexts might need to be co-developed, changing and sometimes instituted (e.g. a lecture)These may need to be different for different groupsSome kinds of social behaviour are necessarily context-dependent (compliance)It is unlikely that a lot of key social knowledge, behaviour etc. will be generic and hence amenable to explicit programmingMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 12

Development of Social ContextThe context-dependency of cognition has a very social use, the development of mutually recognised contextsIf a particular kind of situation is recognisable by participantsThen specific language, habits, behaviours, norms, etc. will start to be developed for that situationThe more that happens, the more the particular situation will be distinguishableOver time, social contexts become institutionalised and easy (for us as observers) to identifyE.g. Lectures, parties, interviews etc.Indeed it seems we construct our world (buildings, etc.) to facilitate our context-dependent intelligenceAffords identifiable opportunities to utilise the knowledge and computation of others, promoting social embedding, which in turn lock people in to these facilitiesMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 13

“General” Intelligence Intelligence: the ability of an actor/agent to decide what to do to achieve its goals in its situation and given its knowledgeThis is involves learning what works in any particular situation where there are time and resource constraints (e.g. how to socially embed when appropriate)Given the No Free Lunch theorems, this means that there is no “best” strategy for all situations, but rather each strategy is suited for particular kinds of situationHumans do not possess a general intelligence, but one with particular biases and aptitudes, e.g. to facilitate social organisationUnlike computation, there is no “general” model of intelligence – any generalisation, meta-strategy, mix of strategies etc. with be a disadvantage in some situationsLooking for a foundational model of intelligence (one that can be specialised for particular circumstances) is hopeless, and recalls the doomed “Hilbert Programme” in mathematicsMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 14

Implications for Models of and Implementations of IntelligenceWhilst there are formal systems (some logics, set theory, Turing Machines etc.) that are “complete” in theoryThey all make some things easier to do and some harder, thus having representational biasesIt does not mean there is any general and effective means of finding the right computation even if we know one existsRather, the structure of any intelligence needs to be suited to its environment and goalsIn particular, the intelligence of a social agent needs to be suited to its social environmentMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 15

Herbert Simon Administrative Behaviour (1947) observed how people actually behaved in context, and it was not in any ideal form of rationality, but rather a set of relatively simple, but specific procedures“The human being striving for rationality and restricted within the limits of his knowledge has developed some working procedures… These procedures consist in assuming that he can isolate from the rest of the world a closed system containing a limited number of variables and a limited range of consequences.” (1967)Not the same as his later, and much copied, concept of “bounded rationality” which is just an hobbled version of an ideal rationalityMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 16

An Illustration of Simon’s (1967) Rational Decision MakingMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 17

Summary – Mundane RationalityThere is no such thing as a general intelligence or rationality, even in theoryAll intelligence is biased towards its environment/problems Human intelligence is biased:Its survival advantage was social and thus so is our intelligenceIt uses a context-dependent heuristic: a mixture of context-recognition and crisp, relatively simple beliefs/strategies within these contextsPatterns of behaviour can be very specific to the particular context, but might be quite simple once the right context is recognisedThese patterns will be highly socially influenced but in a way that is specific to the context (which might include the group)Mundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 18

Consequences for Representing Human Behaviour in our ModelsGiven all this there are three approaches:Make simple specific models only within a particular context (“in vitro” modelling)Identify the separate contexts and have a (at least somewhat) different model for each along with a mechanism for recognising when context switchesAllow agents to individually learn the specific behaviours in a context-dependent manner and socially situated manner (resulting in among other things social embedding)Mundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 19

1. Behaviour within a single contextModels might be simple, but you don’t know a priori which are applicable in the contextIt might be that traditional models of rationality are OK in some contexts (especially when the context is designed for these models as in many experiments)But it might be even simpler than theseAsking participants, observing them, analysing narrative accounts are good places to startDownsides include: difficult to delineate scope of context, specificity of resultsMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 20

2. Multi-context specificationSpecify N contexts, and for each of these a behavioural model (as in last slide)But people may consider the same situation in different mental contexts, e.g. a lecture as a dramatic performance vs. a lecture as a job interviewEliciting when people mentally flip between contexts is hard as this is largely unconsciousBut can be done by presenting examples to participants within contextsAlso can involve an exponential effort in NMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 21

3. Agents individually learning Context and within context behavioursAllow the agents to themselves learn the appropriate contexts along side learning the appropriate beliefs, language, habits, norms etc. within those contextsIn other words to allow the agents to become individually acculturatedRequires a model of context-dependent learning and reasoning etc.It takes time!Difficult for us (as researchers) might find it difficult to understand the culture that results, since we are not acculturatedMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 22

Model StructureHow the:context-identification system (CIS), the context-dependent memory (CDM), the local learning algorithm (LL), and Inference system (IS) could work togetherMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 23

950900Volatility - past 5 periods850800750700750850950Volume - past 5 periodsSnapshot of (most frequent) model domains in one traderMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 24

The (simplified) model contents in snapshot of one traderMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 25

Some Simulation Work addressing Context-Dependency in Simulations(Schlosser & al 2005) argue that reputation is context dependent(Edmonds & Norling 2007) looks at difference that context-dependent learning and reasoning in an artificial stock market(Andrighetto & al 2008) show context-dependent learning of norms is different form a generic method(Tykhonov & al 2008) aregue that trust is context dependentMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 26

ConclusionGeneric models of how decisions are made are doomed, although one can pretend to oneself that the “variance” from a generic model is random (due to different decisions coming about in different individual mental contexts), this itself is an assumption with consequencesOne can (sometimes) use simple models for behaviour, but only within contextA lot of “field work” needs to be done to identify strategies, habits, norms etc. within contextsA need for techniques to systematically and transparently analyse qualitative data to inform our models of within context behaviourMundane Rationality, Bruce Edmonds, NESS@ECCS, Vienna, September 2011, slide 27

The End Bruce Edmondshttp://bruce.edmonds.nameCentre for Policy Modellinghttp://cfpm.org

Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 29Comparison in an Artificial Stock MarketEnvironment:Traders (n context, n not)1 Market maker: prices and deals: 5 stocksEach fundamental (the interest given on owning stocks) changes by a slow random walkTraders buy and sell shares at current market price, but do not have to do so (there is simple cash too)Traders have memories, can observe actions of others, index, etc.Traders have a limited budgetModelling ‘arms-race’Actions change environment

Agent Cognitive ModelA 2D memory space of volatility vs. volume because these are associated with the “feel” of a market Current values of volatility and volume determine a radius of considered models, the current best (evaluated over a past period) is chosen and used to determine next actionCompared to agents with a similar algorithm but a “point” memory space (all models equally accessible)Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 30

Agent modelsRich language of models for their expectations of future pricesIncluding those that enable chartist, fundamentalist, personal imitation, calculations on past prices, averages, own last actions etc.Agents decide what to do, based on projections using their current best model, their budget, transaction costs, fundamental, actions of others etc.Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 31

The primitives allowed in modelsAND, averageDoneByLast, averageIndexOverLast, averageOccuredToStockLast, averageSaidByLast, boundedByMaxPrice, divide, dividendOf, doneByLast, F, greaterThan, IBoughtLastTime, IDidLastTime, indexLag, indexLastTime, indexNow, indexTrendOverLast, ISaidLastTime, ISoldLastTime, lagBoolean, lagNumeric, lastBoolean, lastNumeric, lessThan, maxHistoricalPrice, minus, myMoney, NOT, onAvIBoughtLastTime, onAvISoldLastTime, OR, plus, presentStockOf, priceDownOf, priceLastWeek, priceNow, priceUpOf, randomBoolean, randomIntegerUpTo, randomPrediction, saidByLast, T, times, totalStockValue, volumeLastTime; F, indexLastTime, indexNow, maxHistoricalPrice, myMoney, onAvIBoughtLastTime, onAvISoldLastTime, randomBoolean, randomPrediction, T, totalStockValue, volumeLastTimeplus: the names of the traders, stocks and a random selection of constantsSocial Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 32

Model StructureHow the:context-identification system (CIS), the context-dependent memory (CDM), the local learning algorithm (LL), and Inference system (IS) could work togetherSocial Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 33

A CDM Learning and Update AlgorithmSocial Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 34D3.72.10.92.2Some Space of Characteristicsp

Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 35Total Assets in a Single RunBlack=context, White= non-context

Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 36950900Volatility - past 5 periods850800750700750850950Volume - past 5 periodsSnapshot of model domains in one trader

Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 37The (simplified) model contents in snapshot of one trader

Conclusion of ExampleIn the stock market example, nothing always works better, since other agents adapt to any winning strategy by a fewThe advantages of embedding or not-embedding are mixed herebut contexts: recognised market “moods” did co-develop in certain periodsThere was a complex “lock-in” between the individual abilities and the organisation, one not easily predicted before experimentationTaking inspiration here from an observed system for how to program agents and how they relate for that systemSocial Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 38

About the model1D space for trait (or location) with tolerance following Riolo et al. (2001) Nature, 411:441-4432 essential foodstuffs, each needs learning to extract via matching a pattern with a modelIndividuals adapt their models (one for each foodstuff) that allows them to try and extract each foodstuffShare food and models with those within their toleranceFood if they have excessModel if (a) their model is fitter or (b) they are olderDepending on nutrition levels they can reproduce (with possible mutation) die etc.Two kinds of model – suitable fuzzy model or unsuitable exact expressionDerived from a model of symbiosis/specialisation (Edmonds 2006)Evolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 39

Traits with ToleranceEvolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 40

Exact, not forced to specialise, model acceptance on fitnessEvolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 41

Exact, forced to specialise , model acceptance on fitnessEvolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 42At time 100At time 100

Fuzzy, not forced to specialise, noisy transmission, model acceptance on fitnessEvolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 43

Fuzzy, not forced to specialise, acceptance of models by ageEvolutionary Models of Information Transmission in Human Societies, Bruce Edmonds, Darwin09, UIB, Mallorca, Nov. 2009, slide 44

Second example of learning models to predict heart disease Social Intelligence and Multi-Agent Systems, Bruce Edmonds, MALLOW 2010, Lyon, 31st August, slide 45serum cholesterolresting blood pressureOnly models with more than two instances shown

Mundane Rationality as a basis for modelling and understanding behaviour within specific contexts

More Related Content

Viewers also liked (7)

Similar to Mundane Rationality as a basis for modelling and understanding behaviour within specific contexts (20)

More from Bruce Edmonds (20)

Recently uploaded (20)

Mundane Rationality as a basis for modelling and understanding behaviour within specific contexts

Editor's Notes