Humans have a tendency to invent new problems rather than solve old ones. As we build larger, more complex systems, we unearth global challenges around networks, compute resources and data. Have we neglected to see more elegant examples which existed all along?
It is possible for even the most complex systems to be organized and simplified in ways that may not occur to us. In situations where we still search for the right algorithms, by turning to complex natural systems around us we can find the problem was solved long ago. What we think is a new protocol may in fact be one that has been tested and evolving over hundreds or millions of years. One invented for the early internet is incredibly similar to a strategy evolved by desert ants millions of years ago. And this is why it works.
This talk will address these questions with examples of self-organization, decentralization and diversification from emergent phenomena found in nature.
10. @helenaedelson
“Any living cell carries with it the
experience of a billion years of
experimentation. ”
- Max Delbrück
Protein molecules within a cell (in green) self-organize
in response to stress: www.pierce-antibodies.com
Data Lineage
patterns evolved to solve problems
11. @helenaedelson
– Margaret Wheatley “Three Images” Noetic Science Review (Spring 96)
“We live in a world which in constantly exploring what’s possible,
finding new combinations – not struggling to survive,
but playing, tinkering, to find what’s possible.”
21. @helenaedelson
Our Brains Are
The Ultimate Collective
• Every decision is the outcome of a neural collective
computation
• Nothing in the brain tells the rest of it to think or remember
Neurons fire signals that only collectively create intelligence
22. @helenaedelson
Patterns In Collective
Computation
This pattern of information accumulation and consensus is
seen in neurons, ants and bees, monkey societies, and many
other systems.
• Neurons go out and semi-independently collect information
about the noisy input, like neural crowdsourcing
• Then come together and reach a consensus on what the
decision should be
23. @helenaedelson
Swarm Intelligence
Computational systems inspired by emergent amplification of
collective intelligence, through the cooperation of hundreds to
millions of homogeneous agents in a system.
Applicable anywhere there is collective decision-making, e.g.
search optimization, network routing, image analysis, data
mining, training neural networks, democratic elections and
fluctuating markets.
24. @helenaedelson
Swarm Intelligence
• Autonomy: many agents networked together, interacting locally
• Decentralization: no leader, supervisor or global coordination
• Order: spontaneous self-assembly into emergent patterns
• System-level patterns are unpredictable from behavior of its members
• Group intelligence and capabilities far exceed the individual
complex adaptive systems that behave in unpredictable ways, wholly
different than the behavior of its parts
26. @helenaedelson
No One In Charge
• Autonomy: many agents networked together, interacting
locally
• No leader, supervisor or global coordination
• No leader election or follower
• No single unit in the network knows what’s going on overall
• Nothing tracks or knows all events and change
convergence between neurons, bees and ants
27. @helenaedelson
Autonomous Agent
• Simple instructions and feedback loops
• Subject to common laws (gravity, aerodynamics)
• Common processing environment and perception systems
• Common goals
• Influence and limit each other's actions
autonomous agents don’t exist in pure chaos,
shared principles bind them together
28. @helenaedelson
Network Diffusion and Contagion
How something spreads over hundreds to
thousands of unique nodes (not clones)
• Regular continual interactions and
computation
• Amplification: through the node to node
feedback loop
• Eventual synchronization
amplification across the noisy collective
29. @helenaedelson
Swarm Algorithms
Swarm intelligence algorithms and strategies are distributed,
decentralized, adaptive, scalable and incorporate randomness for
performance.
• Particle Swarm Optimization (PSO)
• Ant Colony Optimization (ACO)
• Artificial Bee Colony (ABC)
• Stochastic Vehicle Routing problem (VRP)
• Traveling Salesman Problem (TSP)
• Bee Nest-Site Selection Scheme (BNSSS)
30. @helenaedelson
Computational Agent
• Limited capabilities and intelligence
• Governed by a set of very simple rules
• Rules provide criteria to make decisions
• Shares information with proximal peers
• Communicates often through brief interactions
individuals behave like neurons in a human brain
32. @helenaedelson
– Adrian Dyer, Researcher, Royal Melbourne Institute of Technology
“Our computers are electricity-guzzling machines.”
The bee, however, “is doing fairly high-level cognitive tasks
with a tiny drop of nectar.
Their brains are probably processing information in a very clever way.”
33. @helenaedelson
Bees
The colony’s collective behavior enables solving complex
tasks like:
• Maintaining a constant temperature in the hive
• Keeping track of changing foraging conditions
• Selecting the best possible nest site
34. @helenaedelson
House Hunting Algorithm
How does a colony solve the life or death problem of finding a new
home? They hold a democratic debate.
• Search: highly distributed searching by scouts
• Assessment: evaluation of potential sites, based on criteria
• Advertise: locally communicate information about a resource
• Consensus: each evaluates for quorum threshold
• Relocation: transport to the new home
colonies and consensus
36. @helenaedelson
– Deborah Gordon, Biologist, Stanford
“Ant algorithms have to be simple, distributed and
scalable – the very qualities that we need in large
engineered distributed systems"
37. @helenaedelson
Chaos Or Pattern?
The world's largest ant colony stretches over 2.7 km2 / 670 acres, contains
approximately 306 million workers and 1 million queens across 45,000
interconnected nests.
38. @helenaedelson
Hundreds of thousands of travelers speed along densely packed highways,
transporting huge loads, without congestion.
Congestion Avoidance Optimization
Inbound
Outbound
Outbound
Army ants have evolved a three-lane traffic system
39. @helenaedelson
Exchanging Information
Ants interact via smell with their antennae, or if it encounters a short-
lived patch of pheromone deposited by another.
With one quick touch, an ant can identify
• A nest-mate - established trust
• What task the other has been doing
identity, trust and task
40. @helenaedelson
One algorithm we work with was invented in the early stages
of the internet because operating costs were high. Its goal
was managing data congestion by gauging bandwidth
availability.
It is incredibly similar to one evolved by desert ants to gauge
resource scarcity, many millions of years ago.
41. @helenaedelson
• When an ant forages in the sun it loses water
• It gets water back from seeds it eats
• Would-be foragers wait at a narrow tunnel
entrance to the nest
• As returning food-bearing foragers pass, they
drop their load to briefly touch antennae with
those waiting (the positive feedback loop)
Foraging Strategy
Harvester ants evolved an algorithm for conserving water in the desert. They have
to spend it to get it.
42. @helenaedelson
The rate of interactions drive decisions of
individuals.
• It doesn't matter which ant it meets
• Only the rate at which it meets other ants
Foraging Optimization
rate of interactions over content
Additionally they had to solve searching for resources that are scattered (by
wind and flooding), with unpredictable spatial dispersal versus predictably
clustered.
43. @helenaedelson
Acks that trigger transmission
of the next data packet and
indicates available bandwidth.
A forager leaves the nest in
response to the rate it meets
returning foragers with food.
TCP Three-Way Handshake
congestion avoidance and determining availability
Just as the rate of packet transmission increases/decreases with the rate of
returned Acks, the rate of outbound foragers increases/decreases with the rate
of successfully returning foragers.
44. @helenaedelson
A source sends out a large
wave of packets at the
beginning of a transmission to
gauge bandwidth
Foraging harvester ants send
out scout foragers to gauge
food availability before auto-
scaling the rate of outgoing
foragers
TCP slow start
gauging bandwidth & elastic scaling
45. @helenaedelson
Timeout when a data transfer
link breaks or is disrupted, and
the source stops sending
packets
When foragers are prevented
from returning to the nest for
more than 20 minutes foragers
stop going out.
TCP Timeout
system stays stopped unless a positive event occurs
49. @helenaedelson
- Radhika Nagpal, Professor of Computer Science,
Harvard University Wyss Institute for Biologically Inspired Engineering
“The beauty of biological systems is that they are elegantly simple,
and yet in large numbers, accomplish the seemingly impossible.
At some level, you no longer even see the individuals;
you just see the collective as an entity to itself.”
50. @helenaedelson
Distributed
Robotics
• A single simple robot has many
limitations, and can only do a few
simple things
• Yet, at scale, the smart algorithm
overcomes its physical and
mathematical limitations
51. @helenaedelson
AI Algorithms At Scale
• Schools of autonomous underwater vehicles coordinating with no
central leadership to
• gather data on ocean currents and ecology
• monitor or clean up pollution
• Hundreds of robots cooperating for quick disaster response
• Millions of self-driving cars on our highways
53. @helenaedelson
Algorithms Tuned By Evolution
• Flexible roles
• Decentralization, No leader
• No reporting to one particular unit
• Distributed consensus
Unus pro omnibus, omnes pro uno
• Simple rules and instructions
• Local interactions and feedback loops
• Self-organizing
• Super-coordinators
With the right organization, a group can solve cognitive problems with an
ability that far exceeds that of its members.
Resilience and Reduced complexity:
54. @helenaedelson
Reliability and Resilience
Resilience is a measure of a system’s ability to survive and persist within
a variable environment.
• Societies like monkeys, swarms or proteins in a cell have evolved
strategies to survive shock
• Swarm networks respond efficiently to attack and disruption through
simple interactions
• These networks are easy to repair and can grow or shrink because
they evolved to tolerate randomness
how to thrive in a random world
55. @helenaedelson
Adaptation and Rapid Exploitation
The capacity of collectives to quickly learn, adapt and invent new patterns is
much higher than top-down command/control.
• In a flood of possibly conflicting neural signals, our brains have to quickly
compute what we perceive and decide how respond
• If ants or bees encounter a roadblock they quickly experiment with options
and rapidly exploit a viable solution (like ensemble forecasting)
a single visual neuron is like a single bee or ant scout
56. @helenaedelson
– Buckminster Fuller
“You never change things by fighting the existing reality.
To change something, build a new model that makes the
existing model obsolete.”
@helenaedelson
58. @helenaedelson
Resources
• Pattern Discovery over Pattern Recognition: A New Way for Computers to See
• The 1000 robot swarm
• Swarm intelligence and neural network for data classification
• Smart swarms of robots seek better algorithms
• Neural Underpinnings of Decision Strategy Selection: A Review and a Theoretical Model
• Collective Computation
• A Markov Chain Algorithm for Compression in Self-Organizing Particle Systems
• The effect of individual variation on the structure and function of interaction networks in harvester ants
• The Remarkable Self-Organization of Ants
• The ants go marching, and manage to avoid traffic jams, Princeton Weekly Bulletin
• The Regulation of Ant Colony Foraging Activity without Spatial Information
• A Survey On Bee Colony Algorithms
• Artificial Bee Colony Optimization: A New Selection Scheme and Its Performance Andrej Aderhold, Konrad Diwold, Alexander
Scheidler, and Martin Middendorf
• How and why trees talk to eachother
• What Is Spacetime
• Chaos Theory, The Butterfly Effect, And The Computer Glitch That Started It All
• Scott Camazine, “Self Organization in Biological Systems”
• Protein aggregation after heat shock is an organized, reversible cellular response
• Chaos, Meaning, and Rabbits: Remembering Walter J. Freeman
• Distributed House-Hunting in Ant Colonies Mohsen Ghaffari Cameron Musco Tsvetomira Radeva Nancy Lynch {ghaffari,
cnmusco, radeva, lynch}@csail.mit.edu, MIT
• Phototactic Supersmarticles
• How Nature Solves Problems Through Computation
• How Ants Use Quorum Sensing To Estimate The Average Quality Of A Fluctuating Resource