Hidden Order: How Adaptation Builds Complexity by John Holland

Summary

Holland walks us through how coherence emerges from unstructured agents in environments of continuous flux; coherence under change and complex adaptive systems (CAS)

Key Takeaways

Behavior depends much more upon interactions of agents than their actions
Catalog of all activities does not equal understanding the effect of changes in the ecosystem
Many complex systems show coherence in face of change through extensive interactions, aggregation of diverse elements and learning/adaptation
1. Must understand the interactions and dynamics of the system before can hope to make any significant, lasting changes
Theory can help detect lever points where small changes lead to big outcomes – Amplifier Effect
1. Cross-disciplinary comparisons are vital as subtle characteristics in one context can be easily drawn out in others
CAS systems made up of a large number of active elements diverse in form and capability
1. Makes system stronger and more robust. Weeding out weak actors so that only the strong remain counter-intuitively leads to worse performance
Rules are used to describe agent’s strategies – few, simple rules can lead to complex behavior
1. A major part of the modeling effort for any CAS, then, goes into selecting and representing stimuli and responses, because the behaviors and strategies of the component agents are determined thereby. Once we specify the range of possible stimuli and the set of allowed responses for a given agent, we have determined the kinds of rules that agent can have
Adaptation – process by which an organism best fits itself to its environment
1. Time scale of adaptation varies drastically and they are very important to take into account in any system (humans vs. trees)
  1. The fast dynamics will shape the slow
2. Overall, we will view CAS as systems composed of interacting agents described in terms of rules. These agents adapt by changing their rules as experience accumulates. In CAS, a major part of the environment of any given adaptive agent consists of other adaptive agents, so that a portion of any agent’s efforts at adaptation is spent adapting to other adaptive agents, co-evolution (Red Queen). This one feature is a major source of the complex temporal patterns that CAS generate
The 7 Basics
1. Aggregation
  1. Simplifies complex systems by grouping similar things which leads to constructing models as these are prime building blocks
  2. Emergence of complex, large-scale behavior from aggregate of small, simpler behaviors (ants and “intelligent” ant colony)
2. Tagging
  1. Facilitates the formation of aggregates as tags manipulate symmetries (flag as a rallying cry which helps group people together)
  2. Tags enable us to observe and act on properties previously unobservable due to symmetries (spinning white cue ball harder to spot but when a stripe is added you can easily tell in which direction it is rotating)
  3. Facilitates selective interaction – filtering, specialization, cooperation leads to emergence of meta-agents and organizations through individual agents are always changing
  4. Tags are the mechanism behind hierarchies
3. Non-linearity
  1. whole is greater than the sum of the parts
  2. Behavior in aggregate more complex than the parts would indicate
4. Flows
  1. Nodes (processors, agents), connectors (designate possible interactions), Resource
  2. Adapt as time elapses and experience accumulates
  3. Tags almost always define the network by delimiting the critical interactions, the major connections. Tags acquire acquire this role because the adaptive processes that modify CAS select for tags that mediate useful interactions and against tags that cause malfunctions
  4. Multiplier Effect – resource injected in one node spreads over network which leads to chain of changes (big in network/flows modeling)
  5. Recycling Effect – the effect of cycles in the network can drastically increase output of the system over time as the system retains resources and these resources can be further exploited as they offer new niches to be exploited by new kinds of agents. This process leads to increasing diversity through increasing recycling (virtuous cycle)
5. Diversity
  1. Each agent fills a niche which is determined based on interactions centering on that agent
  2. Nature abhors vacuums and will fill empty niche with new agent – typically similar in form and habit (the convergence effect, mimicry)
  3. CAS systems get diverse via adaptation which leads to further interactions and new niches – symbiosis, parasitism, mimicry, biological arms races
  4. Perpetual novelty is a hallmark of CAS
6. Internal models
  1. Mechanisms CAS used to anticipate – eliminate details so that selected patterns are emphasized. Agent must select patterns in the torrent of input it receives and then must convert those patterns into change sin its internal structure
    1. A model allows us to infer something about the thing modeled
  2. Tacit and overt models
    1. Tacit simply prescribes a current action, under an implicit prediction of some desired future state
    2. Overt model is used as a basis for explicit, but internal, explorations of alternations, a process often called lookahead
  3. Natural selection selects for better internal models
7. Building blocks
  1. Deconstruct complex problem into simpler parts which can be used and reused in different circumstances
  2. The search for powerful building blocks is the most effective way to make the best internal models
    1. We can a significant advantage when we can reduce the building blocks at one level to interactions and combinations of building blocks at a lower level: the laws at the higher level derive from the laws at the lower level building blocks. This does not mean that the higher level laws are easy to discover but it does add a tremendous interlocking strength to understanding systems and hierarchies
CAS exhibit coherence under change via conditional action and anticipation and do so with no central controller.
Can discover lever points if can uncover general principles which govern CAS dynamics
Agents must act somewhat similarly if a uniform approach to CAS is feasible
Adaption – a rule’s ability to win based on its usefulness int he past – older rules are likely best as they’ve been tested by time
1. Credit Assignment to best rules easiest when have immediate feedback – tests the rule’s utility
  1. Bucket Brigade – the credit assignment procedure which strengthens rules that belong to chains of action terminating in rewards
2. Agent should prefer rules which use more information about a situation
  1. Higher specificity leads to stronger rules (higher in the hierarchy)
3. Default hierarchy – early on, agents will depend on overly general default rules that serve better than random actions. As experience accumulates, these internal models will be modified by adding competing, more specific exception rules. These will interact symbiotically with the default rules and the resulting model is called a default hierarchy. Default hierarchies expand over time from general default to specific exceptions (the young man knows the rules, the old man the exceptions)
4. Adaptation by rule discovery – trial and error may work but doesn’t leverage system experience
  1. Plausibility – take strong rules and apply to new areas which seem promising
    1. Innovation / creativity – simply combining tested building blocks in new ways
5. Recombination of rules leads to discovery and occasionally mutation which can produce a more fit offspring
  1. More fit building blocks are used more frequently which are then passed on more often to succeeding generations
  2. More complicated building blocks usually formed by combining simpler blocks
    1. The higher level are typically composed of well-tested, above-average simpler blocks. Over time, the hierarchy becomes more elaborate, providing for the persistence of still more complex behavior. A
  3. Reproduction, recombination and replacement (genetic algorithm) found in nearly every CAS system
  4. Implicit parallelism – individuals (no matter how great) don’t recur but their building blocks do
    1. Evolution “remembers” combinations of building blocks which increase fitness
  5. Discovery of new building blocks leads to a slew of new innovations (punctuated equilibrium)
With any model, must know what has been emphasized (exaggerated) to make a point and what has been left out to keep focused on that point
Hierarchy – the appearance of new levels of an organization in this evolution depends on one critical ability: each new level must collect and protect resources in a way that outweighs the increased cost of a more complex structure. If the seeded aggregate collects resources rapidly enough to “pay” for the structured complexity, the seed will spread.
Successful approach to any theory – interdisciplinary; computer-based thought experiments, a correspondence principle (models should encompass standard models from prior studies in relevant disciplines); a mathematics of competitive processes based on recombination

What I got out of it

Fascinating book on how the universe seems to produce order for free via coherence, spontaneous self-organization and complex adaptive systems.