LEARNINGCORE

Multi-Armed Bandit

Balance pulling the best-known lever against testing new ones. Regret is the score.

λ mean ≈ variance
counts per window · recreated schematic

How it works

  • Each pull teaches and earns — explore or exploit, every round.
  • Regret (vs the best lever in hindsight) is the scoreboard.
  • Optimism under uncertainty: try what could plausibly be best.
Use → Trials, dosing, ad creative. Run small experiments, then exploit.
Watch out → Declaring a winner after ten pulls.
Try it → A/B/C test with 10% exploration for a month.
REDCAPE jobs → Act · Explore
Free sources → Wikipedia: Multi-armed bandit (CC BY-SA) • Coursera M11 (Learning)
◍ Mentor soonSee REDCAPE playbook →

Attach to a lattice — combines with

Different families, shared jobs. Disagreement between them is where insight lives.

DYNAMICS
Markov Chain
Next state depends only on now. Run it forward to get the long-run mix.
SHARED JOBS: ACT · EXPLORE
DIFFUSION
SIR Contagion
Susceptible–Infected–Recovered with an R0 tipping point. Same math for flu and pop-star fever.
SHARED JOBS: ACT · EXPLORE
GROWTH
Capital Share (r > g)
When returns on capital outrun growth, wealth concentrates without any villain.
SHARED JOB: ACT

More in Learning

Replicator Dynamics (Fisher)
Winning strategies reproduce faster; adaptation speed rises with variation.
Rugged Landscapes (NK)
Interacting parts create many local peaks; hill-climbers get stuck, diverse searchers find higher ground.
Diversity Prediction Theorem
Crowd error equals average error minus diversity. Diverse models beat the lone genius.
← Replicator Dynamics (Fisher)Rugged Landscapes (NK) →