// simulate runs a Monte Carlo simulation with given arms and strategys
func simulate(bs strategys, arms arms, sims, horizon int) (simulations, error) {
ret := simulations{}
for _, b := range bs {
s, err := sim.MonteCarlo(sims, horizon, arms, b)
if err != nil {
return simulations{}, fmt.Errorf(err.Error())
}
ret = append(ret, s)
}
return ret, nil
}
func TestSoftmax(t *testing.T) {
τ := 0.1
sims := 5000
trials := 300
bestArmIndex := 4 // Bernoulli(bestArm)
bestArm := 0.8
arms := []sim.Arm{
bmath.BernRand(0.1),
bmath.BernRand(0.3),
bmath.BernRand(0.2),
bmath.BernRand(0.8),
}
strategy, err := NewSoftmax(len(arms), τ)
if err != nil {
t.Fatalf(err.Error())
}
s, err := sim.MonteCarlo(sims, trials, arms, strategy)
if err != nil {
t.Fatalf(err.Error())
}
expected := sims * trials
if got := len(s.Selected); got != expected {
t.Fatalf("incorrect number of trials: %d", got)
}
accuracies := sim.Accuracy([]int{bestArmIndex})(&s)
if got := accuracies[len(accuracies)-1]; got < 0.9 {
t.Fatalf("accuracy is only %f. %d sims, %d trials", got, sims, trials)
}
performances := sim.Performance(&s)
if got := performances[len(performances)-1]; math.Abs(bestArm-got) > 0.1 {
t.Fatalf("performance converge to %f. is %f", bestArm, got)
}
expectedCumulative := 200.0
cumulatives := sim.Cumulative(&s)
if got := cumulatives[len(cumulatives)-1]; got < expectedCumulative {
t.Fatalf("cumulative performance should be > %f. is %f", expectedCumulative, got)
}
}
func TestDelayedStrategy(t *testing.T) {
τ := 0.1
sims := 5000
trials := 300
flushAfter := 100
bestArmIndex := 4 // Bernoulli(bestArm)
bestArm := 0.8
arms := []sim.Arm{
bmath.BernRand(0.1),
bmath.BernRand(0.3),
bmath.BernRand(0.2),
bmath.BernRand(0.8),
}
b, err := NewSoftmax(len(arms), τ)
if err != nil {
t.Fatalf(err.Error())
}
d := NewSimulatedDelayedStrategy(b, len(arms), flushAfter)
s, err := sim.MonteCarlo(sims, trials, arms, d)
if err != nil {
t.Fatalf(err.Error())
}
accuracies := sim.Accuracy([]int{bestArmIndex})(&s)
if got := accuracies[len(accuracies)-1]; got < 0.9 {
t.Fatalf("accuracy is only %f. %d sims, %d trials", got, sims, trials)
}
performances := sim.Performance(&s)
if got := performances[len(performances)-1]; math.Abs(bestArm-got) > 0.1 {
t.Fatalf("performance converge to %f. is %f", bestArm, got)
}
expectedCumulative := 200.0
cumulatives := sim.Cumulative(&s)
if got := cumulatives[len(cumulatives)-1]; got < expectedCumulative {
t.Fatalf("cumulative performance should be > %f. is %f", expectedCumulative, got)
}
}
func TestSoftmaxGaussian(t *testing.T) {
τ := 0.1
sims := 5000
trials := 300
bestArmIndex := 1 // Gaussian(bestArm)
bestArm := 5000.0
arms := []sim.Arm{
bmath.NormRand(5000, 1), // is five times better
bmath.NormRand(0, 1),
}
strategy, err := NewSoftmax(len(arms), τ)
if err != nil {
t.Fatalf(err.Error())
}
s, err := sim.MonteCarlo(sims, trials, arms, strategy)
if err != nil {
t.Fatalf(err.Error())
}
expected := sims * trials
if got := len(s.Selected); got != expected {
t.Fatalf("incorrect number of trials: %d", got)
}
accuracies := sim.Accuracy([]int{bestArmIndex})(&s)
if got := accuracies[len(accuracies)-1]; got != 1.0 {
t.Fatalf("accuracy is only %f. %d sims, %d trials", got, sims, trials)
}
performances := sim.Performance(&s)
if got := performances[len(performances)-1]; math.Abs(bestArm-got) > 0.1 {
t.Fatalf("performance converge to %f. is %f", bestArm, got)
}
expectedCumulative := 4500.0 * float64(trials) // (mean(bestArm)-tolerance) * num trials
cumulatives := sim.Cumulative(&s)
if got := cumulatives[len(cumulatives)-1]; got < expectedCumulative {
t.Fatalf("cumulative performance should be > %f. is %f", expectedCumulative, got)
}
}