HIDDEN ORDER
Chapter 1: Basic Elements
Complex Adaptive Systems (cas)
• Aggregates of independent agents• Agents behavior governed by a collection
of rules• Persistence and coherence in the face of
change– Dependent on extensive interaction,
aggregation of diverse elements, and adaptation and learning
• i.e. economy, ecosystem, CNS
Objectives
• Purpose of the book is to formulate a solid theory about complex adaptive systems– Separate fundamental characteristics from
idiosyncrasies – Difficult to do because of nonlinearities– Compensate with cross-disciplinary
comparisons
Seven Basics
• Four properties and three mechanisms shared by all cas
• Not the only shared characteristics.
• The rest can be built from a combination of these seven
Aggregation (Property)
• 2 senses• 1st Sense: Constructing Models• 2nd Sense: What cas do
– Emergence of complexity from a network of simpler agents.
– Aggregates can act as agents in higher-level structures (meta-agents)
– i.e. Cells Tissues Organs Organ Systems Organism
Tagging (Mechanism)
• A mechanism that facilitates aggregation by bringing like things together– i.e. a flag used to rally a group of people,
pheromones and visual patterns that lead to selective mating
• Allow observation of properties hidden by symmetry.– i.e. painting a stripe on a cue ball to see
changes in the axis of rotation
Nonlinearity (Property)
• A function is linear its value can be calculated by adding the values of its parts.– i.e. A=B+C
• Nonlinearlity of Pool Ball Example– (B+W)(t+1)=(B+W)+c(B)(W)– (G+W)(t+1)=(G+W)+c’(G)(W)– Can we simplify the model by aggregating B and G into a single
category S? (S=B+G)– No. There is no coefficient that works for all combinations of B
and G. – If the function was linear, then we could take the average of c
and c’
Flow (Property)
• [node, connector, resource]• [phones, phone lines, conversations]• Multiplier Effect
– Occurs after additional resource is added at a node– Initial effect is multiplied as the resource is passed
through the network
• Recycling Effect– Result of cycles in a network– Recycling with the same input produces a greater
output at each node
Diversity (Property)
• Not random.
• System will adapt to fill holes.– New agent typically occupies same niche as
previous agent and provides most of the missing connections
– Leads to convergence in biology• i.e. squid vs. mammalian eye
– Founder Effect
Diversity (Property)
• Pattern of diversity in cas is dynamic.
• Diversity is a product of adaptation. Each new adaptation creates possibilities of new interactions and new niches
Internal Models (mechanism)
• Allows for anticipation
• Two types– Tactit: prescribes an action under an implicit
prediction of a desired future state• i.e. a bacteria swimming up a nutrient gradient
– Overt: used as a basis for explicit explorations of alternatives (lookahead)
• i.e. exploring possible scenarios before moving a chess piece
Building Blocks
• How are internal models built from constantly changing environments?– Situations are distilled into useful and relevant
building blocks– A flat tire while driving a red Saab on the
expressway?– Decompose into rules about cars, tires, and
expressways
Hidden Order Chapter 2
Adaptive Agents
By John Holland
Requirements of Adaptive Agents
1. Performance System - a universal way to represent the capabilities of different agents using messages and conditional rules
2. Rule Discovery - making changes to an agent’s capabilities by means of a credit score
Making Use of Binary
• A detector receives a message when conditions are met, the result of which triggers an action and completes the rule
• Messages - strings of 1’s and 0’s of length (L)
• Conditions - strings 1’s, 0’s and #’s (signifying no specificity), also of length L
Messages from Two Sources
1) Detector Originated - derived from and given meaning by the environment
2) Rule Originated - produced by other rules, given meaning only when they activate effectors
Credit Scores
• Rules have strength that is like having cash on hand, and can be viewed as being producers, middlemen, or consumers that buy, sell, and trade messages
• The ultimate consumer is the rule that is active when the agent receives a reward from the environment
• All rules down the line are credited and automatically strengthened
Building Blocks
• Rules can be divided into building blocks called schemata.
• Successful rules have schemata that in general serve as better building blocks than random binary strings
• Successful schemata are singled out by replacing other parts of the sequence with the * symbol
Crossing Over
• Binary sequences are crossed over using the genetic algorithm
• Short schemata are conserved through probability
• Longer schemata are preserved because they are composed of shorter successful schemata that appear in a large portion of the gene pool
Mutation
• Successful rules can become fixed in a population before other more successful rules can be discovered
• Random mutations disrupt fixation during crossing over, restoring adaptation
The Prisoner’s Dilemma
• A game where two players simultaneously decide to defect (D), or cooperate (C)
• Tit-for-tat is a favorable strategy that produces a relatively large number of points
• Agents that begin with schemata for tit-for-tat discover better strategies by deciding on whether an opponent is bluffing or not