Skip to main content

Model Library

Evidence So Far tracks how confident the site is in each result. This page is the complementary map: every simulation model on the site, organized by family, with what each one tests and a link to go read it. Use it to find a specific model rather than to judge how conclusive it is.

How to Read This Page

Each table has an Adaptation column that says what kind of process the model runs:

  • Evolution — selection on an inherited trait or strategy across generations, with no within-lifetime learning.
  • Learning — within-lifetime adaptation only, with no evolutionary component.
  • Evolution + Learning — both act together, which is the site's central two-timescale claim.

Nowak Mechanisms

Five thin wrappers over a shared Moran-process engine, one per classical mechanism for the evolution of cooperation identified in Nowak (2006).

ModelAdaptationWhat it demonstratesExplore
Kin SelectionEvolutionCooperation spreads when benefit routed to relatives, weighted by relatedness, exceeds its private cost — Hamilton's rule rB > C.Kin Selection
Direct ReciprocityEvolutionCooperation sustained by repeated encounters: a cooperator can reward or punish the same partner next round, but only if they meet again.Direct Reciprocity
Indirect ReciprocityEvolutionCooperation sustained by reputation: agents help others whose public reputation signals they are good cooperators, without needing repeated encounters.Indirect Reciprocity
Network ReciprocityEvolutionCooperation sustained by spatial structure: local-neighborhood interaction lets cooperator clusters form and shield each other from exploitation.Network Reciprocity
Group SelectionEvolutionA second level of selection on top of individual competition: the most successful group is periodically copied into the least successful group.Group Selection
Display 1: The five Nowak mechanisms, all built on the shared Interaction Kernel routing engine.

Ecological Models

Four case studies that embed a cooperation problem inside a spatial, ecological setting rather than an abstract Moran process.

ModelAdaptationWhat it demonstratesExplore
Spatial AltruismEvolutionThe minimal case study: whether an altruistic inherited type can survive local exploitation once reproduction is local and empty space is a competitor.Spatial Altruism
Cooperative HuntingEvolutionWhether an inherited, continuously-valued hunt-investment trait can spread when it costs the individual predator but pays off only through successful shared hunts.Cooperative Hunting
Spatial Prisoner's DilemmaEvolutionWhat inherited response rules spread when agents repeatedly play a local Prisoner's Dilemma and reproduce into nearby empty cells using accumulated energy.Spatial Prisoner's Dilemma
Retained BenefitEvolutionThe most abstract case study: strips away mechanism-specific structure to ask how much of the value cooperation creates has to be routed back to cooperators for it to spread.Retained Benefit
Display 2: Ecological evolved-cooperation case studies, each with a browser replay.

Learning × Selection Interaction Models

The models that test the site's central claim directly: that within-lifetime learning and between-generation selection can work better together than either alone.

ModelAdaptationWhat it demonstratesExplore
PredPreyGrassEvolution + LearningA richer predator-prey-grass ecology used to test whether the genome needs a direct channel into individual behavior for selection-driven behavioral evolution to become visible.PredPreyGrass
Model 1: Trust LearningEvolution + LearningBaseline two-timescale model: agents learn a scalar trust value per partner within a lifetime, and evolution selects the inherited parameters that shape that learning.Model 1
Model 2: Q-learningEvolution + LearningSame two-timescale framework as Model 1, but with action-value (Q-learning) reinforcement learning replacing scalar trust.Model 2
Model 3: ExtendedEvolution + LearningAdds reputation, partner choice, and forgiveness on top of Q-learning, shifting the outcome toward conditional cooperation with active monitoring.Model 3
Network Diversity ExperimentEvolution + LearningNot a new model — runs Models 1 through 3 under systematically varied network conditions to test when partner-memory alone stops being enough.Network Diversity
Display 3: The two-timescale simulation family plus PredPreyGrass, this site's largest evolution-and-learning case studies.

Learned Cooperation Models

The Prisoner's Dilemma and Repeated Prisoner's Dilemma pages set up the game-theoretic background; the PPO Study is the model that actually runs it.

ModelAdaptationWhat it demonstratesExplore
PPO StudyLearningTwo independently trained PPO agents repeatedly play a Prisoner's Dilemma, testing whether learned behavior converges to persistent defection, cooperative conventions, or mixed horizon-dependent patterns.PPO Study
Display 4: The learned-cooperation model, with no evolutionary component.

Foundational Replications

From-scratch replications of the papers that established the two-timescale claim this site builds on.

ModelAdaptationWhat it demonstratesExplore
Ackley & Littman (1991)Evolution + LearningA neural-network-controlled creature that senses, moves, eats, fights, and reproduces, testing whether combining an evolving genome with within-lifetime reinforcement learning beats either alone. This site's strongest confirmed result.Ackley & Littman (1991)
Hinton & Nowlan (1987)Evolution + LearningThe needle-in-a-haystack landscape that gave the Baldwin effect its modern computational footing: whether within-lifetime learning can manufacture a gradient where genetic search alone has none. Includes an interactive demo.Hinton & Nowlan (1987)
Prosser (2022)Evolution + LearningTests what changes when learning acts on genetically correlated groups of traits, with the correlation structure itself evolving, instead of on independent single loci.Prosser (2022)
Display 5: Replications this site's two-timescale claim is built on. See The Baldwin Effect for the underlying concept.