Techniques to Support Collective Intelligence on the Web: towards a global brain

 Francis Heylighen
Principia Cybernetica Project
 
 

Collective Intelligence

 Intelligence = problem-solving capacity

 increases
with the number of problems that can be solved
with quality of solutions

Collective intelligence

 synergy
group can find more/better solutions than the sum of solutions found by individuals

Examples

 working groups, organizations
football teams
complex adaptive systems

  • "dumb" components, "intelligent" whole

  • hive mind
  • ant trails, termite mounds, bee dances, ...
  • Obstacles to collective intelligence

     cognitive limits of individuals
    complexity of coordination
    noisy communication
    power games
    establish "pecking order"
    ignore contributions of lower ranking individuals

    Problem-Solving

     Problem = difference between present state and goal

     Problem representation:

     search space: set of all possible states

    action: changes one state into another

    preference function: to select best action for state
    depends on goal and domain knowledge

    "mental map": structured as network

  • states -> nodes
  • actions -> links
  • preferences -> weights attached to links
  • Problem-solving

     find (shortest) path from present state to goal
    can be improved by:
    more actions/more states
    better selection

    How can we extend this framework to collective problem-solving?

     Coordination without Communication

     Stigmergy

     external sign that incites individuals to work on it
    e.g. heaps of mud incite other termites to add to the heap

    Collective action requires a shared "object"

     e.g. ball in football game
    object is focus for action

    External memory

     state of the object registers state of the problem
    actions by different individuals change the state
    results of actions are accumulated in a "memory trace"

    Examples

     notes, documents
    shared "blackboard" or "workspace"
    pheromone trails left by ants

    Shared Mental Maps

     Stigmergic object + Mental map => Shared Mental Map (SMM)

     shared access

     all agents can locally "read" the SMM
    (all) agents can locally "write" to the SMM
    therefore, SMM must be external to the agents

    represents collective domain knowledge

     functions as external memory
    has more knowledge than any individual

    supports collective problem-solving

     keeps track of what has already been achieved
    shows possible alternatives (actions)
    helps decision-making by suggesting best trajectory

    Dynamics of Collective Problem-solving

     agents can contribute to the SMM in different ways-

    from simplest to more complex:

     Averaging

     individual preferences are linearly added
    robust and reliable because of law of great numbers
    no direct interaction
    e.g. democracy, N. Johnson simulation

    Discussion

     individuals make suggestions and give arguments
    allows improving on democratic decisions
    e.g. meetings, GDSS, CSCW

    Self-organization

     non-linear interaction with pos. and neg. feedback
    weak signals can be amplified
    noise can be eliminated
    e.g. ant trails

    Division of labor

     different individuals contribute to different subproblems
    specialization allows greater expertise

    Workflow

     subproblems are automatically directed to most competent individuals
    "active" SMM

    Market mechanisms

     price functions as stigmergic signal to attract "suppliers"
    competition selects for most competent contributors
    disadvantages: positive feedback can get out of hand, loss of diversity through monopoly formation

    From DIS to Global Brain

     DIS provide perfect support for a worldwide SMM

     perfect, unlimited memory
    accessible by everyone
    easy to change

    Traditional DIS lack the needed intelligence

     no active guidance of problem-solving
    little support for dynamics of cooperation

    New techniques can add all necessary functions

     support active, collaborative problem-solving
    make DIS self-organizing
    from external memory to external "brain"

    The "Global Brain" metaphor can help us design a better DIS

     implementing most of the methods discussed

    Stages in Global Brain Development

     Metasystem Transitions

     emergence of higher level of complexity/control

    One-to-one communication

     simple reflex <-> telephone, telegraph

    One-to-many communication

     mass media, TV, radio

    Many-to-many communication

    complex reflex <-> DIS

    Associative Learning

    synaptic learning <-> creation of links

    Combining non-linked information

     thought <-> DIS agents

    Creating new knowledge

     creative thought <-> knowledge discovery

    The Learning Web

     Until now, the Web only "learns" through its users

     poor linking patterns
    too complex for efficient coordination

    Associative Web learning

     Hebb rule: concepts encountered together are linked more strongly
    Adaptive hypertext learning rules:
    if A->B->C, then:

    frequency: strengthen A->B
    transitivity: strengthen A->C
    symmetry: strengthen B->A

    These local rules lead to global self-organization (cf. Bollen)

    Web learning creates links and preferences

     link strength proportional to usage
    averaging of individual preference with pos. feedback
    possibility for "suggestions" (e.g. annotations)

    Collaborative Filtering

     Associations implicit in patterns of co-occurrence

     Examples

  • Amazon bookshop suggests related books based on what other clients bought
  • Firefly agent suggests CD's on basis of lists of preferences

  • People with similar interests are expected to prefer similar things
    => individual preferences determine collaborative selections

    Co-occurrence data are available in many places

     indexes of links to pages on same subject
    pages that are linked in different places
    individual bookmarks, buying patterns, web caches, etc.

    Data can be used to create links and preferences

     probability of co-occurrence P(A|B) determines association strength

    Search Agents

     The Web can actively support problem-solving

     user formulates problem ("query")
    program or "agent" searches through web for solutions

    Search engines

     user enters keywords
    engine finds documents with these terms
    requires global index
    documents must contain exact keywords
    not very selective

    The Thinking Web

     Spreading activation

     application of local search
    key documents are "activated"
    activation follows links to connected documents

  • proportional to link strength

  • new documents get sum of incoming activations
    can find documents without the exact terms

    Agents can find experts

     e.g. from expressed interests
    use "outside consultants"
    allows "workflow"

    Agents as "thoughts" in the global brain

     spread by following associations
    locally combine distributed information

    Solving Complex Problems

     Ill-Structured Problems

     only vague associations
    e.g. diarrhoea, constipation, cramps, colon, gas, bloating...
    instead of "How to cure Irritable Bowel Syndrome?"
    spreading activation finds all related documents
    including documents on Irritable Bowel Syndrome
    may need several iterations

    Logical Queries

     XML standard allows the use of ontologies on the web
    different types of links and nodes: ISA, HAS_PART, ...
    can turn web into semantic network (cf. Joslyn)
    allows "intelligent" inferences
    e.g. "Which birds cannot fly?"
    activation restricted to the right type of links

    Knowledge Discovery

     "Serendipitous" agents

     cf. presentation of A. Riegler
    roam the web to find unexpected regularities
    "data mining" with unlimited database

    Discover general rules

     apply different machine learning techniques
    e.g. most instances of property A have also property B
    -> A implies B

    Discover new concepts

     clustering of similar ideas to produce new categories
    produce new documents by integrating existing ideas
    e.g. indexes

    Conclusion

     Collective intelligence requires Shared Mental Map

     can be represented as network of nodes, links and preferences
    individuals should be able to read and write the SMM
    let collective knowledge emerge
    SMM should actively guide problem-solving

    Techniques to make the Web into a better SMM:

     learning links from usage
    inferring links from co-occurrence
    using agents to search the network
    search engines
    spreading activation
    semantic inferences
    locating experts
    using agents to discover new knowledge

    Thus, the Web can support a true collective intelligence: a "Global Brain"