Background¶

My undergraduate degrees are in mathematics and human geography. My professional experience is in software development for large-scale data visualisation. My interests are algorithmic music, open-ended creative systems, the economics of sustainability and technological innovation, and in general questions of multi-agent learning. “Collective dynamics”, if you like. The flux of hidden orders that constrain the chaos.

Research interests, broad-brush¶

Dynamics of collective learning, with an emphasis of what this means at the economic end of the spectrum.

What kind of stochastic process is multi-agent learning in human beings? How can we model it? Artificial chemistry? Wolpert-style COINs? Feldman-style “Turing gas”? Pattern-matching string transformations? Information theoretic bounds on agent model updating? A network representation of hypothesis formation? For what its worth, my intuition is that something like a combination of artificial chemistry and a statistical model of pattern matching could give some insights into a toy model

If the aggregate system is, as seems likely, unpredictable, what bounds does knowing about an underlying stochastic process place on system evolution? How different would technology be if we “ran the tape twice”? Is there an underlying topology to what innovations can be fostered? Surely the limits imposed by every individual agent’s learning imposes certain constraints on what overall structures can be evolved.

If the adaptive process of innovation is constrained by the structure of adaptive human learning, how is it constrained by the underlying physical reality? Reality, viewed from the perspective of making and testing hypotheses about it, is not a homogeneous state space with constant reward, but must possess a fitness landscape that favours some combination of truth and ease of applicability. (Solar panels work best in the desert, anti-gravity machines don’t work anywhere, Newton’s equations of motion are more readily deducable than relativistic ones at the velocities at which we commonly operate, ready availability of fossil fuels favours polymer-based construction materials etc - do you get my drift?) Can we capture that in some sense?

The attraction of thinking at this degree of abstraction is that it seems like a contender to remove the computational intractability of agent-based models - we can hope that processes of innovation are approximately static, that processes of diffusion are approximately like the standard contagion-on-a-network models, that learning rates by something like a reinforcement-learning convergence rate, and that we can get a model that will show us how variable and weird actual innovation processes are, and how abstract information flows are constrained by reality, but avoid having to simulate decision processes for millions of human agents. Of course it would be an odd system, dominated by black-swan events, innovations that are unlikely for any one individual to make but then rapidly propagate. Average over the whole earth, no individual is likely to invent the transistor, but a global economy that fosters the possibility of such events will crack it sooner or later, say. Closer to Paul Ormerod’s stochastic shocks than Ising’s spin glasses.

On that latter, physical technology note, I suspect that the fitness landscape of economic innovation would have something to do with informational constraints of human learners, but also of the geophysical landscape - what sources of free energy are available? How readily may these be transmitted? There is a lot of work with this in material stocks and flows analysis, and also in the field of ecology. It seems this is related to your “free energy” link, though a more common term and slightly different term from the engineering/ecology literature is “exergy” - thermodynamically available energy. (Approximately - how much energy Laplace’s demon can on average extract from a system with noise.) The free energy thing seems to rest a bit to heavily upon equilibrium hypotheses for my taste, but this underlying computational mechanics idea that a combination of available energy plus information about how likely it is to be available is one of the underlying drivers for learning/replicating systems, that seems reasonable.

Interaction of adaptors on different time-scales - evolutionary versus cultural time-scales, dynamics that are hard or easy to learn, frequent and infrequent event-types... can any regularities survive such heterogeneity? Convergence theorems applied to nonstationary targets.

So, questions this approach seems useful for are: what are the transitions paths to non-carbon-intensive energy systems? How can we quantify the “disruptiveness” of a technology? Can we identify unfilled technological niches in this way? What would a society based on alternative energy forms look like? Which industries are dead in the water?

Research interest, miniaturist’s detailing brush¶

What is a good research étude toward these large issues?

Potential data sets¶

I’m short of data with which to explore problems. Welcome to the ubiquitous ICT epoch though, I can definitely get access to some of those classic data sets online to bootstrap my enquiries.

Economics is so successful in part because it is a self-quantifying system. Prices are already, definitionally in commensurable units. What brilliance, to launder so much complexity through such a narrow bottleneck. Anyway...

• Twitter firehose

• Patent citation networks (these are available and reasonably well annotated)

• MMOG play data might be available and is an amazing source of information here.

• Wikipedia articles and their references (readily available)

  • also includes easily-parseable mathematical data and theorems
  • ...and edit trails
  • ...and category annotations
  • and semantic metadata

• source code of large collaborative projects (Linux or BSD kernel, openoffice, python, Perl, GCC etc)

  • can I parse such projects to see how interfaces form?
  • Are there odd stylised facts about contribution to these that I might be able to explain?
  • Or call-graphs?
  • Will this do anything more than rediscover Dunbar-like numbers?

  This is possibly a low-hanging fruit for me - I’ve got a fair bit of experience of parsing, and SCM-wrangling. But is software engineering a good proxy for physical engineering? Can I parse other technical standards in the same way? (I know a few engineers - must ask them)

• Brand registration records. (Thinking of Holland-style tagging hypotheses here)

• Journal article cross-references. This is quite well-studied.

• Estimating number of SKUs as a surrogate for divisions of a modern economy a la Beinhocker (lots of research into this because of Long Tail theories, though the primary data is rarely included - might chase this.)

• Bank or other firm lifetimes

• Stocks and flows analyses of the economy (These are big in the ecological economics literature, once again, without primary data usually)

  • What is the state of the art in Material Flow Analysis?
  • Can we at all get at the cascades of innovation that occur around scarcity, presuming such innovation does happen in actuality?

• We have Census records also

• ...and public data providers such as infochimps.

What else is digitised? Consumer purchase data, traffic flow data?

How about - what small subsystems in here are documented well?

Groysberg’s Chasing Stars model benefitted from an industry which goes to the trouble of collecting data on itself for him. Do we have any equivalents in areas I am actually passionate about?

Related things¶

• Economics versus biology
• The unfortunate uselessness of most state-of-the-art-academic-monetary-economics
• Coming of Age for Multi-Agent System Research?
• Economists’ Grail: A Post-Crash Model