Methods
A method is how decisions are made. Each round it is given the state (scores, credits, levels still in delivery, the round) and returns a decision: one or more actions, each a group, a mode and a chunk of income, with the reason it was chosen. An experiment names one method; everything else about it (nation, position, scenario, objective, horizon, seeds) is held fixed, so methods can be compared on the same draws.
Every method
Built methods have experiments behind them; the tree search is exploratory; the evolution strategy is named by a brief but not yet built.
- Greedybuilt
greedy
Each round, the actions with the most value per credit.
- Mirror a peerbuilt
mirror-peer
Each round, closes the largest gaps to a peer nation's 2026 profile, by the cheapest open mode.
- Randombuilt
random
Five actions a round, each a random open group and mode and a random chunk of income.
- Holdbuilt
hold
No actions. Income pays maintenance; whatever is left is carried over, losing 5 per cent a round.
- Scriptedbuilt
scripted
Actions fixed in advance by the experiment: for a backcast, the nation's historical pathway.
- Tree search (MCTS)exploratory
mcts
Each round, a Monte Carlo tree search over candidate decisions, played out to the horizon, commits the decision with the most visits.
- Evolution strategy (CMA-ES)exploratory
cmaes
An evolution strategy over a whole plan, scored by playing it through the engine under shared shock seeds: open-loop, or re-optimised every round.