We build an agent-based model (ABM) of how senior politicians navigate the complex governance cycle using relatively simple heuristics. They first test whether they can form a single party minority government. If not, they seek coalition partners and negotiate with these. They treat “Gamson’s Law” – government parties get perks payoffs in proportion to their seat shares – as common knowledge. When different politicians attach different importance to the same issue, "logrolling" allows them to realize gains from trade and agree a joint policy position even when they have divergent policy preferences. We allow for the realistic possibility that multiple proposals for government are under consideration at the same time. Nonetheless, there may often be a “Condorcet winner” among the set of proposals, which beats all others in pairwise comparisons. Finally, we specify a model of government survival, which assumes incumbent governments are subject to a stream of unbiased random shocks which may perturb model parameters so much that legislators now prefer some alternative to the incumbent. For any given government, our model allows us to estimate the probability of this happening.

We set out the case for computational social science as opposed to traditional “pencil and paper” formal methods. The substantive theme of this book is the governance cycle in parliamentary democracies, but the ideas we put forward can be applied to many other areas of study.

We first calibrate and then analyze our ABM using suites of Monte Carlo simulations, applied to a representative set of training cases of government formation in European parliamentary democracies. For each to the twenty training cases, we execute 1,000 model runs, randomizing model parameters for each run as follows. For each observable parameter, for each model run for each training case, we take the empirically observed value and perturb this with parameterized random noise. For unobservable model parameters, we randomly sample from the full range of possible values. The 1,000 runs for each case thus yield a distribution of model-predicted outcome for that case. We calibrate unobservable model parameters by selecting ranges of these associated with empirically accurate model predictions. We analyze the (calibrated and uncalibrated) model by summarizing the mapping of model inputs into model outputs in the artificial data generated by the set of Monte Carlos, using theoretically informed logistic regressions. This is the computational analogue of analyses based on deductive “comparative statics” generated by traditional formal theorists.

We set out an alternative, “top down”, approach to agent-based modeling. We develop an artificial intelligence (AI) algorithm to navigate the governance cycle using what we can think of as computational game theory. AI models have had formidable success in solving games like Chess, Go, and especially a bluffing game like Poker, suggesting they also have the potential to attack difficult political games. Addressing a simplified version of the government formation process as a noncooperative game, the AI algorithm deploys Monte Carlo Counterfactual Regret (MCCFR). During in massively repeated self-play, it samples paths though the vast game tree to relentlessly learn near optimal strategies.

We describe the institutional environment for the governance cycle in parliamentary democracies and the preferences of senior politicians over key political payoffs. We are not concerned here with electoral politics, so treat an election as a “black box” which, in expectation, administers unbiased random shock to party seat shares. Elections trigger government formation. The government, once formed is subject to a steam of unbiased shocks, some of which may perturb either the environment or the preferences of senior politicians sufficiently to cause them now to prefer some alternative to the incumbent government. The more susceptible an incumbent to such shocks, according to the model, the less stable it is likely to be. Politicians’ policy preferences are described in terms of their ideal positions on a large number of binary issues, and the relative importance they attach to each issue. The utility they derive from any government is described as a convex combination of the distance between their policy preferences and the agreed government policy position, which may involve “agreeing to disagree” on some issues; and their share of the fixed perks of office.

While heavy-duty computational methods have revolutionized much empirical work in political science, computational analysis has yet to have much any impact on theoretical accounts of politics – in contrast to the situation in many of the natural sciences. We set here out to map a path forward in computational social science. Analyzing the complex and deductively intractable “governance cycle” that plays out in the high-dimensional issue spaces of parliamentary systems, we use two different computational approaches. One models functionally rational politicians who deploy rules of thumb to navigate their complex environment. The other deploys an artificial intelligence algorithm which systematic learns, from massively repeated self-play, to find near-optimal strategies. Future work made possible by greater computational firepower would enable better AI, more realistic ABMs, and the modeling of logrolling under the conditions of incomplete information which characterize most real-world bargaining and negotiation.

We use our calibrated ABM and our AI algorithm to make case-by-case predictions of outcomes in new out-of-sample test data. These predictions concern: the full partisan composition of the cabinets which form, participation by particular parties in the cabinets which form, and the observed durations of the cabinet which forms. Absent a baseline model of government formation in such complex settings against which we can evaluate our results, we compare success rates with those of a prediction of minimal winning coalitions which is common to a large number of existing studies. Bearing in mind that the ABM in particular generates probability distributions of predicted outcomes in each case, which we feel is substantively realistic, while only a single outcome can be observed, we are very satisfied with the predictive accuracy of the model. Successful predictions relating to cabinet durations are particularly distinctive to the model, deriving from the model-predicted number of issues tabled in formation negotiations, and the model-predicted likelihood than a random shock will create a situation in which a majority of legislators now prefer some alternative to the incumbent.

We outline the core argument of the book and steps taken to establish this. We begin by sketching component parts of the governance cycle: election, government formation, and government survival. Noting that the analysis of this complex system is intractable for traditions deductive methods of formal modeling, we preview two different computational methods for analyzing it. First, we model “functionally rational” artificial agents who use simple but effective rules of thumbs to navigate their high stakes but complex environment (ABM). Second, we specify an artificial intelligence (AI) algorithm which, by massively repeated self-play, teaches itself to find near-optimal strategies for playing what is in effect a traditional, but intractable, noncooperative game. We conclude by sketching the empirical approach we use to first calibrate and exercise the models on training data and then test them on out-of-sample test data.