Controls
Run simulation
Animate
Reset time
draw 0
Ready.
Definitions
agent chose optimal action
agent chose suboptimal action
group row success
group row failure
Model
K states / K actions : world q makes action q uniquely optimal.
Payoff variance : target variance among action success probabilities within each world.
Outcome : selected action yields binary payoff 0/1 with its world-relative success probability.
Strategy presets
Best-up-to-now : choose an action with maximal observed average payoff; ties are random.
Decreasing ε-greedy : explore randomly with probability 1/(t+1); otherwise choose best-up-to-now.
Social/fast ε-greedy : explore randomly with ε = max(floor, 1/(observed evidence+1)^α); otherwise choose best-up-to-now.
Preferred-action family : each agent has preferred action i mod K; with probability 1/(t+1), choose it; otherwise choose best-up-to-now.
Reinforcement learning : choose action a with probability w(a)/Σw; add observed payoffs to the corresponding action weights.
Metric definitions
Mean : average of all agent correctness cells at time t.
Coverage : fraction of histories where at least one agent chose the optimal action.
Jθ : fraction of histories where the proportion of correct agents is at least θ.
Alignment : average pairwise correlation of agents' correctness variables across histories.
SV : average within-history variance across agents.
MV : variance across histories of the column mean.