Procedural Models of Political Order
The point of a model is not to replicate real-world phenomena; it is to identify which inputs are most important in producing a given output. The simplest functional model tells us what are the necessary and sufficient inputs to achieve the output we are interested in. This, in turn, tells us where we should focus our attention in real life — which input changes will have greatest effect on the output and which are peripheral.
However important a factor may seem at street level, if it can be removed from a model without any significant effect on the accuracy of the output then it is unimportant. Take the homo oeconomicus model of economic transactions. It is often criticised on the basis that it implies rational decision-making, even though we all know that humans seldom make entirely rational decisions. As a result of this, many have argued that we should abandon conventional macroeconomic models and replace them with ones that factor in imperfect rationality. Why, then, has this not already been done?
Because, as it turns out, incidences of individual irrationality have no effect whatsoever on macro-level market phenomena, and attempting to work them into the models tends to make them worse rather than better. Important as irrational decision-making may seem in day-to-day life, it is simply not a relevant variable at the aggregate level, where seemingly irrational individual choices come together to produce a larger emergent rationality, with the thought processes of the market as a whole being independent from and qualitatively different to those of any individual participant. Under such conditions, an individual is to the market as a neuron is to the brain: a single neuron cannot make reasoned decisions, but a billion can, and the macro output will look nothing like the micro outputs that went into producing it. However, it is extremely difficult to guess the emergent properties of a whole simply by observing the component parts. This is why modelling is important, and why parsimony is an important criteria for judging any model: the most parsimonious model tells us what — among a million and one possibly-relevant phenomena — is important to the output, and what is not.
From Model to Algorithm to Visualisation
In a previous article we put forward two extremely parsimonious models designed to describe the way in which political power is structured in Anglo-European and Chinese-influenced societies.
In Western political philosophy, the principal threat to human life comes from other humans. The development of political order is an attempt to mitigate this threat. Under such a system, the roles of leader and follower are defined by means of a simple utilitarian calculus: each individual identifies the person that constitutes the greatest threat to him and associates himself with whoever seems able to provide the most effective protection against this threat. Thus, unaffiliated weaker individuals will tend to follow the second most powerful individual in the system at any given time, as a way of hedging against the most powerful actor. The result is continual turnover at the leadership level: an aspiring leader promises protection from the greatest present threat, attracts followers, achieves dominance, gradually comes to be seen as a threat himself, and is replaced in his turn.
The mechanism can be observed in action in the checks-and-balances systems and multiparty democratic arrangements that still prevail among Anglo-European polities. Under this system no one group can remain in power for long, as citizens gradually come to perceive them as having grown overweening and transfer their allegiance to whichever opposition group seems capable of throwing the rascals out.
While this fractious, high-turnover approach to politics frequently appears chaotic, it is in fact the key to stability within those states in which it prevails, preventing the long-run dominance of any particular group. The precise pretext for political conflict at any given moment — left vs. right, rich vs. poor, centre vs. periphery — is unimportant; what matters is merely that it be perpetual. The existence of this unending conflict gives ordinary citizens the ability to shop around between different sources of protection. Thus they are able to keep costs and exploitation to a minimum, by threatening to take their custom to a competitor should membership of their present faction grow too onerous.
The result is a series of nested protection rackets: my party protects me against the opposition, my local representatives protect me against the national elites, and my country protects me against foreign threats. Just as in any protection racket, power is an all-or-nothing game. If the shops in a particular village are paying the ‘Ndrangheta for protection, it is on the assumption that it will prevent other rival groups from posing a threat. Similarly, a state’s authority depends on it being the only authority on its purported territory. The state’s ability to protect its citizens is the sine qua non of its existence; a state unable to prevent other individuals or entities from violating its laws on its own territory loses that territory. This mechanism was theorised by Max Weber as the “monopoly of legitimate violence” and underpins modern international law.
Because the ideas described above are so prevalent and so efficient in describing political thought in the Anglo-European sphere, it can be tempting to see in them a universal model.
In fact, this Chinese system is not merely older, but has historically enjoyed far greater prevalence. It is not unreasonable to regard the Chinese model as the default pattern of political organisation, and the Western version merely an intriguing alternative. In regions traditionally influenced by Chinese political thought, theories concerning the nature of power and the origins of the state have traditionally been based upon Spring and Autumn and Warring States era ideas of human development, which differ significantly from their Western counterparts. Rather than focusing on human threats, the earliest Chinese narratives of state formation emphasise natural risks, notably floods and famines. In these stories, the first states grew out of the incorporation of communities around individuals who had succeeded in developing new techniques in agriculture and flood defence and were willing to share their knowledge. The earliest sovereigns were innovators in farming and hydrology, whose political legitimacy was based on these skills, rather than upon their ability to protect partisans from human threats. They were described as having attracted followers through their technical inventions, with the followers submitting to their rule in exchange for the better livelihood that proximity offered. Just as in the Western model, the decision to sacrifice independence in order to be part of a community that provides significant quality-of-life benefits was a simple, utilitarian calculation.
Under such a system, a leader is the individual who can render the greatest number of people dependent upon the advantages that he can provide, and threats to his power come not from rival offers of protection, but from redistribution networks that escape his control. Thus, the defining quality of statehood is not the monopoly of legitimate violence, but the monopoly of legitimate benevolence.
As a result, we were accused more than once of oversimplifying things. To test our theories, therefore, we have built two agent-based models, one representing each society. The aim was to judge whether the inputs that we identified were indeed necessary and sufficient to create the structures we described.
To visualise the models, we converted the theories into two algorithms. In both cases agents were randomly assigned a degree of economic surplus and a generosity propensity score. In the Chinese model unaffiliated agents were instructed to search for and follow the agent with the highest surplus*generosity score in each round, with the aim of maximising their profits. In the Anglo-European model they were instructed to search for and follow the agent with the second-highest surplus*generosity score, with the aim of hedging against the most powerful agent. In every round an agent’s surplus*generosity score is recalculated based on the original sum, his total number of followers (since followers bring in revenue as well as demanding redistribution), an economic growth coefficient set by the user, and a random component (suprlus modifier rate) intended to stand in for exogenous benefits and disasters (a flood, an unexpectedly good harvest etc.). A more technical description of the algorithms can be found here.
These were then visualised via a front-end produced using Mesa:
The results were much more interesting than anticipated. Not only did the model produce the structures we predicted, but the outputs reflected emergent properties of the systems that we did not explicitly tell them to simulate, providing strong additional evidence in favour of our overall hypotheses. Thus, for example, while we made no effort to write in any instructions that would force the models to move towards an equilibrium state of ever-longer periods of stasis, they did this consistently. Similarly, we did not explicitly program the agents in the Chinese simulation to respond to adverse economic circumstances by forming more stable networks faster than they otherwise would. Nevertheless, they did just this — something that was, in fact, predicted by the same texts upon which we based our design.
In the following sections we will describe the results obtained from running both models, beginning with the Chinese one on account of its lower degree of algorithmic complexity.
Modeling Chinese Visions of Political Order
A recording of a typical run of the Chinese model using default settings can be found below. The darker the shade of a dot, the more followers it has. Thus the darkest magenta dot is the leader of the “pink faction”, paler pink dots are his lieutenants (who also have followers of their own), and the palest dots are pure followers.
As predicted, the Chinese model tends to produce a single flattish pyramid, with one very strong player, one or two layers of subordinates, and everyone else being relatively equal. When any particular subordinates begin to rise to prominence, this implies that they have succeeded in creating semi-independent redistribution networks of their own, and the leader is liable to be overthrown. (When all or most of the dots in a network change colour at once, this is a sign that a leader has been overthrown by one of his own deputies, who has then assumed control over all of his former leader’s followers.) Thus, the only secure leader is one who is alone, directing from far above the fates of a relatively egalitarian group of followers. In other words, the classical authors’ warnings against factionalism and independent wealth were not merely the product of envy or paranoia. A leader must aim to maximise inequality between himself and his followers while also minimising it between individual followers, because this reduces the number of his followers capable of challenging him for power.
What is most interesting in this model, however, is not these results per se, but the fact that none of our agents are capable of this reasoning. They are programmed simply to redistribute whatever their redistribution propensity coefficient tells them to redistribute and to follow whoever has the highest surplus*generosity score. Nevertheless, the system as a whole behaves as though they were following the precise chain of reasoning described above. This implies that the reasoning is a product of the system rather than vice versa; a system in which the most generous individual predominates is a stable one, therefore the simulation will gravitate towards this equilibrium, not as a matter of conscious choice, but simply as a reflection of the fact that unstable systems collapse quickly and stable ones survive longer. Agents whose decision-making algorithms are rewarded by stable configurations will prosper and impose their strategies upon others — even when no thought at all has gone into the process. Their conscious intellectual reasoning is of no importance; all that matters is their actions - the process is an entirely Darwinian one: every short-term failure eliminates another sub-optimal process and pushes the ensemble a little closer to long-term success, just as the death of every suboptimal individual improves the overall fitness of a species.
(Out of interest, here is the same simulation run under much harsher economic circumstances. As described above, the centralisation process is much faster and less volatile.)
Modeling Anglo-European Visions of Political Order
Being more familiar with the Anglo-European political model, we anticipated few surprises when we made it. We were confident that it would produce a relatively stable two- or three-party system. It did, but not in the way we expected.
We expected to see, in most iterations of the model, the formation of two or three different parties (coded in different colours) of roughly equal strength, which would then compete for followers and hence for power, as unaffiliated agents hedged against the strongest individual in the system by following the second strongest.
What we did not expect was for the two blocs competing for followers and power to be evenly-matched factions of a single party. In the recording - as in most iterations - you will see that the simulation is often dominated by a single colour, with two or three darkly-shaded power centres (in contrast with the Chinese version above, in which alternative power-centres are much paler in colour). While one of these is the “official” leader, the other factional leaders — his deputies — are close to being his equals.
The model, it turns out, produces a far subtler and more nuanced portrait of the strategic compromises and ambiguities inherent in democratic politics than we imagined during the design process. Indeed, analysing the outputs obliged us to re-evaluate many of our own preconceptions regarding the nature of party politics.
If you have ever felt that the established left and right wing parties are merely the same group of people wearing different rosettes, the model would seem to agree. While, in this model, the competition for followers and the possibility for followers to “kick the rascals out” provides a democratic safeguard against tyranny, constraining as it does the leaders’ freedom of action, this cannot be described as a purely competitive system either.
Indeed, if we adopt a purely Darwinian perspective, it makes sense for a group seeking power under such constraints to split itself in two to satisfy the system’s requirement for constant competition while blocking the path to outsiders, precisely as the model suggests. The real world implications of the observed trends are immediately apparent. The same systemic influences explain the ferocious joint opposition on the part of both sides to a relatively independent candidate such as Donald Trump, and his popularity among ordinary voters (who would probably benefit more from genuine opposition than rule by a single bifurcated elite). Similarly, it helps to show why — independent though Trump was — the easiest way for him to rise was via an existing faction, rather than by creating one of his own from scratch. In other words, the model produces a depiction of the race to the centre that is astonishingly and cynically accurate.
It is important to emphasise that we are not suggesting that this dominance by a single elite disguised as two parties is a calculated conspiracy. As we noted before, our agents are simply not capable of such reasoning. These sophisticated schemes for maximising individual and group utility emerge as a pure product of evolutionary pressure. If the system favours the survival of those agents who adopt particular behaviours, the number and influence of agents adopting such behaviours will increase, however the agents do or do not rationalise the process. Just as the computer agents do not make any rational strategic choice, neither do the majority of committed members of the Democratic or Republican parties, who would never declare that they see themselves as belonging to a single bifurcated power bloc. Instead, they will say — with all possible sincerity — that they are on a given side because their opinions align with that side’s beliefs but they are willing to compromise to defeat “extremists”.
The system, for its part, does not care what the agents argue about, as long as they argue. It is this constant low-intensity factional conflict that, gyroscope-like, maintains overall stability, and serves the best interests of both leaders and followers, with or without their conscious awareness. The precise issues at stake at a given time do not matter, they are merely the bait used by the system to incentivise agents to work to ensure its own continued survival. All the op-eds, the Thanksgiving dinner angst, the political philosophy papers, your own profoundly held beliefs… None of them matter; we can simulate it all in a Kaggle notebook and get the exact same result. You may believe you are a complex individual and a sophisticated political thinker, but consider this: we didn’t even need to pay for extra GPU capacity.
The reality is that we argue because the system requires us to argue, and the mechanisms by which it pushes us to do so have evolved following the same Darwinian processes as described in the analysis of the Chinese model above. Those structures that incentivised agents to maintain this equilibrium between factions were selected for, while those that did not were eliminated. Systems that did not attain equilibrium collapsed and were replaced with better ones.
This, in turn, raises some interesting questions regarding intelligence, both artificial and otherwise. These models would not meet most people’s definition of artificial intelligence, and yet to any observer they appear to be learning, and moreover learning to solve relatively difficult problems with sophisticated multi-stage reasoning. Over time they move away from instability to create longer and longer periods of equilibrium — something that many high-IQ humans struggle to manage. We know that this is a product of attrition rather than calculation, but does it matter?
Additional General Observations
Possibly the most interesting practical implication of this exercise has to do with what it has to say about the nature of power and leadership, whether under Anglo-European or Chinese constraints.
Both models suggest that political structures are defined by followers rather than leaders. Our algorithms programmed only follower behaviour, saying nothing explicit about leadership. Instead, by elevating those individuals with the qualities that they required, follower-agents created their own leaders. In our model various levels of willingness to become a leader are implied by the redistribution function proper to each agent, however, it is the followers who have the final say. In each case they are given a particular vision of what a leader is, and they structure their society in such a way as to select the individual that best fits these criteria.
This shows most clearly in the models’ responses to failure. The (Chinese) authors whose descriptions inspired the present models describe a system that fails upwards, with stingy or incompetent leaders’ patronage networks collapsing and being replaced by new ones formed by smarter and more generous individuals. However much disruption the system suffers, as long as more than one individual survives, it will rebuild itself along the exact same lines as before because the survivors are followers programmed to select their own leaders according to one and the same blueprint. A polity may thus be compared to the cells of a live sponge: passed through a sieve, they will immediately begin rebuilding themselves into a sponge. They lack the instructions to do anything else.
This, in turn, explains why attempts to overlay foreign political structures on top of existing ones tend to result in either collapse and corruption (as has frequently occurred in Africa) or a thinly-disguised replica of the old system (ditto in Asia). The functioning of a polity is determined by the programming of its follower-agents rather than by regulations imposed from on high. Attempts to change this will — if effective — create chaos and — if ineffective — a Potemkin village concealing a copy of whatever system existed before the change. Thus modern China is effectively an imperial bureaucracy cosplaying a Marxist-Leninist state, while contemporary Japan is a feudal system in which the daimyo’s ritual duties include occasional election campaigning. This is not an expression of cynicism on the part of the leaders, but a response to demand on the part of the followers.
This is not to say that it is impossible to make a conscious and successful effort to modify the ways in which a given system functions, however. In both the Anglo-European and Chinese systems, relatively small intellectual elites successfully guided a transition from feudalism towards meritocracy by reprogramming agents’ views of how politics should work - in the first case during the Enlightenment and in the second during the Warring States era. However, in each case this effort succeeded because its proponents met two important criteria:
In both cases the change was advantageous for both the system and the agents. The former regained the stability it had lost with the growth of an increasingly entrepreneurial bourgeoisie, while the latter gained greater opportunities than they had previously enjoyed. If the changes had benefited only the system and not the agents or vice versa, the change would have failed.
In both cases it was done via the skillful inflection of pre-existing tendencies — changing the criterion for wealth redistribution from proximity to utility in the Chinese case, and the fulcrum of loyalty from a local to a national one in the Anglo-European case. The definition of leadership did not change — it was still based upon redistribution in the Chinese system and on protection in the Anglo-European one. The only change was in the channels through which this was expressed. Once the new channels were established, agents assimilated them so completely that they were passed down to subsequent generations with no questioning or suggestion that alternatives could be possible or desirable. If the change requires a complete rewiring of agents’ preferences, then it will likely be bypassed rather than assimilated — something explains the prevalence of one-party states across the Chinese cultural sphere, even in nations such as Japan and Singapore that possess entirely democratic institutions.
Models
If you want to play with the models yourself, they can be found here:
https://www.kaggle.com/tsangchungshu/europolitics?kernelSessionId=75854639&kernelSessionId=75854639
If you have never used Kaggle before, you will need to create an account to run the model as it requires an internet connection. First click on “copy and edit”:
Then click on “run all”:
Finally click on the ngrok.io url that will be generated to open the model visualisation:
Post Scriptum
Due to lack of funds, these models sat around without visualisations for quite some time while we worked on other things. Notably, we began looking at variations on the Chinese system as described in classical literature, and particularly at one modified version that incorporates system-level learning in response to external stimuli.
Contributors: Joseph Lim, Prad Nelluru, plus one collaborator who requested not to be named.