Smarter
Parts Make Collective Systems Too Stubborn
By Jordana
Cepelewicz
https://www.quantamagazine.org/smarter-parts-make-collective-systems-too-stubborn-20190226/?utm_source=Quanta+Magazine&utm_campaign=c6b69a1ba3-RSS_Daily_Biology&utm_medium=email&utm_term=0_f0cb61321c-c6b69a1ba3-389846569&mc_cid=c6b69a1ba3&mc_eid=61275b7d81
The
poet John Donne immortalized the idea that “no man is an island,” but neither,
it turns out, are most other species. Many natural and artificial systems are
characterized by the collective, from neurons firing in sync and immune cells
banding together, to schools of fish and flocks of birds moving in harmony, to
new business models and robot designs operating in the absence of a single
leader. “Collective systems are more the rule than the exception,” said Albert Kao, a researcher at the Santa Fe
Institute who models decision-making in such systems.
Whether biological,
technological, economic or social, collective systems are often considered to
be “decentralized,” meaning that they lack a main control hub for coordinating
their individual components. Instead, control is distributed among the components,
which make their own decisions based on local information; complex
behaviors arise through their interactions. That kind of setup can be
advantageous, in part because it is resilient: If one part isn’t working
properly, the system can continue to function — in stark contrast to when a
central brain or leader stops doing its job.
Decentralization has ridden a
wave of hype,
particularly among those hoping to revolutionize marketplaces with blockchain
technology and societies with more dispersed governments. “Some of this
stems from political ideology having to do with a preference for bottom-up
governing styles and systems with natural checks on the emergence of
inequality,” Jessica Flack, an
evolutionary biologist and complexity scientist at the Santa Fe Institute,
wrote in an email. “And some of it stems from engineering biases … that are
based on the assumption these types of structures are more robust, less
exploitable.”
But “most of this discussion,”
she added, “is naive.” The line between centralization and decentralization is often
blurry, and deep questions about the flow and aggregation of information in
these networks persist. Even the most basic and intuitive assumptions about
them need more scrutiny, because emerging evidence suggests that making
networks bigger and making their parts more sophisticated doesn’t always
translate to better overall performance.
In a paper published earlier this month in Science
Advances, for instance, a team led by Neil Johnson, now a physicist at George
Washington University, demonstrated that a decentralized model performed best
under Goldilocks conditions, when its parts were neither too simple nor too
capable. That finding echoes other results from complexity research about the
optimal use of information and the tradeoff between independence and
correlation. The new insights could help to point out the strengths and
limitations of decentralized designs for robots, self-driving vehicles, medical
treatments and corporate structures — and might even help to explain aspects of
natural evolution.
From the Marketplace to the Lab
Johnson’s investigation began as
an attempt to understand the feedback loops in financial systems: Traders, each
trying to maximize their profits while obeying certain rules, make decisions
that contribute to a general outcome — say, some change in stock price — which
in turn influences the traders’ subsequent decisions.
Then one day, while he was still
on the faculty at the University of Miami, Johnson’s attention was caught by
the work of a colleague toiling over something seemingly unrelated to Wall
Street: the movements of fly larvae. A larva crawls automatically to positions
that are neither too hot nor too cold, but it doesn’t rely on its brain to
guide this journey. Instead, each of its body segments responds to signals from
temperature-sensing neurons by compressing on one side or the other. The
collective movement of all the segments causes the larva to turn. The resulting
trajectories toward heat sources reminded Johnson of the financial models he
worked with — so he decided to use them to search for principles common to all
decentralized systems.
He and his team built a model
that mimicked how the larvae behaved. Like the larva’s segments, the collection
of agents in the model shared a common goal but had no way to communicate and
coordinate their activities. Each agent chose repeatedly to move either left or
right, based on whether past decisions had led the entire system to head closer
to or farther away from a specified target. To guide its decision, each agent
used a strategy drawn from a subset of possibilities assigned to it. If a given
strategy performed well, the agent continued to use it; otherwise, it turned to
a different one in its arsenal.
The researchers observed that
when the agents could remember only one or two outcomes, fewer strategies were
possible, so more agents responded in the same way. But because the agents’
actions were then too correlated, the collective movement in the model took it
along a zigzagging route that involved many more steps than necessary to reach
the target. Conversely, when the agents remembered seven or more past outcomes,
they became too uncorrelated: They tended to stick with the same strategy for
more rounds, treating a short string of recent negative outcomes as an
exception rather than a trend. The model became less agile and more “stubborn,”
according to Johnson.
In
an experiment, the movements of decentralized systems hinged on the independent
decisions of their components, which had memories of some past outcomes.
After the memory size reached a certain threshold, the systems performed worse.
The trajectories were most
efficient when the length of the agents’ memory was somewhere in the middle:
for about five past events. This number grew slightly as the number of agents
increased, but no matter how many agents the model used, there was always a
sweet spot — an upper limit on how good their memory could get before the
system started to perform poorly.
“It’s counterintuitive,”
said Pedro Manrique, a postdoctoral
associate at the University of Miami and a co-author of the Science
Advances paper. “You would think that improving the sophistication
level of the parts, in this case the memory, would improve and improve and
improve the performance of the organism as a whole.”
A Second Wave
Kao sees a striking connection
between Johnson’s and Manrique’s findings and his own work on the behavior of
crowds. Over the past few years, he and others have found that medium-size
groups of animals or humans are optimal for decision-making. That conclusion
runs contrary to the standard beliefs about the “wisdom of crowds,” Kao said,
“where the larger the group, the better the collective performance.”
Success lies in achieving the right balance between coordination and independence
among the system’s components.
“It’s like a second wave of this
kind of research,” Kao said. “The first wave was naive enthusiasm
for these collective systems. Now, it feels like … we’re questioning a lot
of the assumptions we made initially and finding more complex behaviors.”
Further research is needed on how
the sophistication of components, their interconnectivity and other parameters
affect the overall robustness and limitations of a network. Johnson and others
plan to study how the availability of information affects phenomena as diverse
as the formation of opinions among voters, the behavior of better robots, and
potential mechanisms of recovery from neurological diseases. They also hope the
work could help explain why natural evolution has made organisms a mix of
centralized and decentralized systems (these kinds of results, Johnson said,
could help to justify “why we aren’t just fantastic larvae”).
Ultimately, complexity in nature
isn’t simply a story of its emergence from entirely uncorrelated groups of dumb
parts — and probing that story could one day yield more universal principles
about cooperation, coordination and collective information processing.
Posts
