← all models

How these models are built — and what that costs

Agent-based models are easy to build and easy to over-claim for. The thesis these come from spends a chapter on the methodology before it runs anything, and the arguments are still the reason to look at the models rather than at the pictures they make. Several of them have a model you can run attached.

An agent's day is one unit of time, and that is the budget

Every Person in these models has one day and spends it. In the earlier models it buys a wage, which buys goods; in the transmission belt it buys goods directly, and the goods are the wage. That is not a simplification for tidiness. It is the answer to a problem the thesis ran into and could not get around any other way.

Model money as a real conserved quantity that agents hand to each other, and it develops dynamics of its own that swamp whatever you were studying — severe inequality out of nothing but a random walk, with no preferences, no strategy and no production anywhere in the model. Using time removes the problem by removing the stock: a day cannot be hoarded.

F1 — zero-intelligence money exchange: the failure that prompted it. Watch the Gini coefficient climb from zero on a model with nothing in it.

Three kinds of decision, and only one of them is dangerous

The framework distinguishes decisions by when they fire, and the distinction is not bookkeeping — it decides which parts of a model can be trusted.

The permutation gap

Agents act one at a time and the order is reshuffled every day, because otherwise whoever goes first has a permanent advantage. Reshuffling is right. But now ask how much time passes between one agent's consecutive actions: go last today and first tomorrow and almost none has; go first today and last tomorrow and nearly two days have. The average is exactly one day and it is almost never one day for anybody in particular.

So every feedback decision in the model is quietly dividing by a denominator that varies by a factor of two. A Firm that moves, waits a turn and asks "did my revenue rise?" is comparing across an interval it does not know the length of. It is a real source of the noise those Firms are misreading, and it is invisible unless you go looking.

F2 — the permutation gap: ten thousand agents, and the distribution is a pyramid centred on one day rather than a spike at it. The checkbox turns on the fix.

How many agents is a modelling assumption, not a detail

Agent counts are usually chosen for speed and reported as if they were incidental. They are not. Several results in these models change with population, and one of them changes in a way that is worth stating carefully.

In the monocentric model, People trade crowding against distance until nobody can improve by moving, and utility ends up equal everywhere. Except at the edge: past a cutoff there is no crowding left to trade away, so those People pay pure distance cost and their utility falls. They are not out of equilibrium — no move improves them — they are outside the range over which equalisation is possible.

That fringe is 44% of the population at 125 People and 0.45% at 2,000. So "utility is flat" is a large-numbers idealisation, exactly true only in a limit, and a model run at 125 People to keep it responsive would report something qualitatively different from the same model at 2,000.

C1 — the monocentric result: switch on the equalising cutoff to see the ring, and the swept figure below it for the population series.

Rational agents can do worse than blind ones

A production function with a fixed labour requirement — output is (l − α)/β, so nothing is produced until α units of labour arrive — is perfectly sound as algebra, where l is an aggregate and nobody has to be first. Put it in a model where agents each supply one unit of labour and it has a hole in it exactly where the agents are.

An agent who reasons correctly works out that joining alone yields less than nothing, and declines. Every agent reasons correctly, so nobody ever joins, and production never starts. An agent who does not reason at all — pick somewhere at random, stay if things improved — eventually stumbles into a crowd, and production starts.

This is why every model here uses a smooth production function instead. It produces something for any positive input, so there is no threshold to coordinate across, and both kinds of agent work under it.

F3 — blind against rational choice: two populations side by side, one of which never produces anything. A3 shows the same function's negative region as a curve.

Showing something is the start of an explanation, not the end

This is the sharpest methodological claim in the thesis, and it is aimed at agent-based modelling generally.

A model that produces a striking result has demonstrated that some mechanism can produce it. It has not said which one, and "emergence" is not a mechanism.

The thesis holds itself to this. Two Firms and one Person, with no space anywhere in the model, produce a catastrophic shift in production structure as one parameter crosses a threshold — a good result, and on its own an unexplained one. So the next model strips out the agents entirely and leaves four pairs of numbers and one line of arithmetic. It does the same thing at the same threshold.

That locates the mechanism exactly. Both Firms are updated simultaneously from the same state, so it cannot be turn order, or sampling, or anything about agents. The only channel between them is that both prices appear in the same demand equation — which is what the utility function is — and one parameter sets how strongly. The instability lives inside the Person's preferences.

Compare that with the money-exchange model, which is the same situation handled the other way: a striking output, a resemblance to real data, and no claim to the mechanism at all. The thesis' footnote on it is the best line in the chapter.

E3 — the regime shift, then E4 — tracking down the transmission belt. Run them in that order.

Functional and descriptive mapping

There are two things a model element can be doing. It can describe something — this agent is a firm, this distance is kilometres — or it can function as something: this term makes crowding expensive, and whether it is rent, congestion or noise is left open.

The density cost in these models is deliberately functional. It is not land rent; there are no landlords, no bid-rent curves and no market for space at all. It is only "being near others costs something", and it is paid by the crowd to nobody. That it produces von Thünen's rings anyway is the result — the rings do not require the institution usually invoked to explain them.

The same discipline is why the thesis keeps love of variety and economies of scale apart when the standard model collapses them into one parameter. Two distinct things doing two distinct jobs, and their interaction is the finding.

C6 — von Thünen without landlords, and C6c, where the sorting reverses because the cost is spatial. G1 — the Core Model is the comparator that cannot separate them.

What this rebuild adds

These models were rebuilt twice from the 2013 originals — once in Java to establish what they do, and once in JavaScript to run here — and the rebuilding turned things up. They are on the model pages under "what we found", including where they complicate the thesis rather than confirming it.

A rebuild that quietly reproduced everything would be less informative than one that does not.