Skip to main content

Ecology, Learning, and Evolution in PredPreyGrass

Morphological, Behavioral, and Life-History Perspectives​

This document is an extensive, self-contained synthesis of the discussion about what kind of evolution is (and is not) present in the current PredPreyGrass environment.
It is intended to be uploaded directly to a repository, documentation site, or shared as a conceptual reference.

The goal is classification and clarity, not prescription.

1. What is meant by Darwinian evolution?​

Darwinian evolution requires three ingredients:

  1. Variation
  2. Inheritance
  3. Differential survival and reproduction

Whenever these are present, selection operates.
The key question is what kind of traits vary and are inherited.

2. The three axes of Darwinian evolution​

Darwinian evolution is most cleanly decomposed into three orthogonal axes:

  1. Morphological evolution — what an organism is
  2. Behavioral evolution — what an organism does
  3. Life-history evolution — how effort is allocated over time

This triad is standard in evolutionary biology and maps cleanly onto artificial agent systems.

3. Morphological evolution​

3.1 Definition​

Morphological evolution is selection on phenotype-defining parameters:

  • body size and shape
  • movement capabilities
  • metabolic efficiency
  • sensing organs
  • physical constraints

These traits define the state-transition dynamics of the organism in its environment.

3.2 Morphological evolution in PredPreyGrass​

In PredPreyGrass, morphology corresponds to body parameters, such as:

  • movement speed and movement cost
  • energy decay and energy gain
  • sensing radius and observation structure
  • action affordances
  • interaction rules (e.g., kill thresholds)

When you define multiple predator or prey types with different such parameters:

  • types compete
  • population ratios change
  • types can dominate or go extinct

This is genuine Darwinian morphological evolution.

4. Behavioral evolution​

4.1 Definition​

Behavioral evolution is selection on strategy, holding morphology fixed.

In biological terms:

  • foraging strategies
  • mating strategies
  • cooperation vs defection
  • social organization

In artificial agents:

  • policies
  • decision rules
  • action-selection strategies

4.2 Behavior vs learning​

A crucial distinction:

  • Learning changes behavior within a lifetime
  • Behavioral evolution changes which behaviors are inherited

Learning alone does not constitute behavioral evolution unless behavior is heritable.

4.3 Behavioral evolution in PredPreyGrass​

In the current setup:

  • policies are shared per agent type
  • PPO updates policies continuously
  • offspring immediately use the shared policy

Therefore:

  • behavior adapts via learning
  • behavior does not evolve via inheritance
  • no strategy can go extinct as a strategy

This is ecology with learning, not behavioral evolution.

5. Life-history evolution​

5.1 Definition​

Life-history evolution concerns traits that regulate timing and allocation, not immediate action:

  • age at first reproduction
  • lifespan
  • reproduction frequency
  • offspring investment
  • semelparity vs iteroparity

These traits are often parameters, not decisions.

5.2 Life-history traits in PredPreyGrass​

In PredPreyGrass, life-history parameters include:

  • reproduction energy threshold
  • reproduction cost
  • reproduction cooldown
  • aging rate
  • starvation dynamics

These traits:

  • strongly affect fitness
  • are heritable per type
  • are not under policy control

They therefore constitute life-history traits, not behaviors.

6. What the current model does evolutionarily​

Putting the above together, the current PredPreyGrass environment implements:

  • ✔ ecological dynamics
  • ✔ death and reproduction
  • ✔ demographic selection
  • ✔ morphological evolution (between types)
  • ✔ within-lifetime learning

But it does not implement:

  • ✘ heritable behavioral variation
  • ✘ selection between strategies
  • ✘ policy extinction or fixation

The correct classification is:

Ecological population dynamics with learning and morphological selection

7. Why reproduction timing is not a strategy​

In the current environment, reproduction follows a rule of the form:

IF energy ≥ threshold AND cooldown passed:
    reproduce()

Key consequences:

  • reproduction is not an action
  • the policy cannot accept or refuse reproduction
  • timing is determined by state, not choice

Thus:

Reproduction is a mechanism, not a behavioral strategy

Policies can only influence reproduction indirectly by affecting survival and energy.

8. Contrast with Genetic Algorithms and PBT​

8.1 GA / PBT​

In GA and Population-Based Training:

  • the population consists of policies
  • policies reproduce and are replaced
  • inheritance is explicit
  • mutation introduces variation
  • poor strategies are removed

Selection acts directly on strategies.

8.2 PredPreyGrass​

In PredPreyGrass:

  • the population consists of agents (bodies)
  • policies are shared
  • reproduction does not transmit strategies
  • death removes individuals, not strategies

Therefore:

Selection removes bodies, not behaviors

9. The minimal boundary to policy evolution​

The system crosses into behavioral (policy) evolution if and only if:

  1. Multiple policy variants exist within the same morphology
  2. Policy identity is heritable
  3. Differential reproduction changes policy frequencies

Mutation is optional; inheritance and selection are sufficient.

10. Clean diagnostic question​

A single question distinguishes ecology-with-learning from behavioral evolution:

Can two agents with the same body but different policies compete, reproduce, and change their relative frequencies over time?

  • No → ecology + learning
  • Yes → behavioral evolution

11. Final synthesis​

Darwinian evolution is not a single monolithic process.

It can act on:

  • Morphology (what organisms are)
  • Behavior (what organisms do)
  • Life-history traits (how organisms allocate effort over time)

Your current PredPreyGrass model already captures:

  • morphology-based selection
  • ecological dynamics
  • learning-driven behavioral adaptation

Behavioral evolution requires only one additional conceptual ingredient: heritable strategy variation within a fixed morphology.

Final takeaway​

Not all evolution is behavioral, and not all behavior is evolutionary.
PredPreyGrass currently implements ecological and morphological evolution with learning; behavioral evolution begins when strategies themselves become heritable and selectable.