Raw results¶
Full results are available on Zenodo:
Online: 'https://sake.re/20260504-DOTP'
Repo: 'https://felapton.inrialpes.fr/cakes/20260504-DOTP.git'
This notebook reproduces, in a self-contained and runnable form, the experiment behind the thesis "Factors Influencing the Evolution of Culture in Population Encounters"
The structure mirrors the scientific argument:
Evaluates the influence of population size, network connectivity, and training set overlap on assimilation or hybridization of behavior and knowledge
Date: 2026-03-20 (Edoardo Cecchini (UGA))
Hypothesis: Larger population assimilates smaller ones; lower connectivity leads to hybridization; larger training set overlap increazes hybridization
10 runs; 150000 games
Experimental setting: Two populations stabilize independently on the same environment and then interact together
This experiment asks: what happens when two populations, each having independently developed its own culture, meet and have to live together? Two outcomes are conceivable:
The goal is to find out which of these occurs, and what influence it.
Agents represent the world with a formal ontology and use it to make a decision about each object they encounter. Through repeated interaction they adapt: whenever two agents disagree, the less successful one adjusts its ontology toward the more successful one (failure-driven adaptation). A population that has reached internal agreement on every object is said to have a stable culture.
The experiment extends this baseline with a two-phase protocol:
Three variables are manipulated to see what shapes the outcome:
Fixed parameters:
Variable grid:
Explored as a full factorial design.
Because an agent's knowledge lives on two levels, the outcome is measured with two distances:
Geometric classification scheme. Let $A$ and $B$ be the two pre-encounter cultures and $C$ the merged one. Place a sphere of radius $\rho = \tfrac{d(A,B)}{2}$ on each of $A$ and $B$. Then:
This is applied both behaviorally (using $d_d$ with $h_1,h_2,h'$) and semantically (using $d_o$ with the ontology sets). The scheme is intuitive but discrete; where it proves too coarse, the analysis falls back on a continuous distance-difference metric introduced where it is first needed.
Full results are available on Zenodo:
The remaining sections load the run's outputs and then work through the four hypotheses, one at a time. Each section states the prediction, shows the result, and interprets it.
First, the end-of-run metrics and the post-encounter interaction logs are loaded, and each run is labelled by its $(\alpha, \eta, \omega)$ configuration.
Each run is then classified independently under the geometric scheme — once on the behavioral distance and once on the ontology distance. Classifying per run (rather than averaging distances first) matters: the scheme's radius depends on the pre-encounter distance of that run, so averaging across runs would produce a threshold never actually observed.
Prediction. As long as the two groups can interact, repairing disagreements should push the merged population to a single behavioral consensus, the cumulative success rate should climb toward 1.
The success rate is cumulative ($\tfrac{\text{successful games}}{\text{games played}}$), so it trends upward as the population converges.
Because the success rate carries the weight of every early failed game, it approaches but never quite reaches 1 within the fixed budget. A cleaner check is the final consensus: the average share of agents agreeing on the majority decision for each object at the end of the run.
Finding: H1 is supported. Almost every run reaches total final consensus (mean ≈ 0.9996); the few runs that stop just short reflect the fixed iteration budget meeting the simulator's randomness, not a failure of convergence. The success-rate curve is a good picture of the trajectory; the final-consensus value is the precise verdict.
Prediction. With skewed sizes, minority agents are drawn into cross-cultural interactions more often and have fewer peers to reinforce their knowledge, so the majority's behavior should prevail.
Below, runs are grouped by population distribution $\alpha$ and counted by outcome (behavioral, then semantic), with a $\chi^2$ test and Cramér's V effect size. As stated before, by convention, for $\alpha < 0.5$ Population 1 is the minority.
Reading the tables. Behaviorally, assimilation by the majority (Population 2) rises from a near-even split at $\alpha=0.50$ to near-total dominance ($\approx 99\%$) at $\alpha=0.12$, a highly significant shift with a medium-high effect size. The semantic tables are dominated by "hybridization" almost everywhere: this is an artifact, not a finding. Ontologies in this model only ever grow (classes split, never merge), so post-encounter ontologies sit far from both baselines and land outside both spheres by construction. This is exactly the limitation that motivates the continuous metric below.
To get past the three-bin scheme, define the normalized distance difference
$$\Delta d = \frac{d(A,C) - d(B,C)}{d(A,B)}$$
so that:
Values near $\pm 1$ mean near-complete dominance.
Normalising by $d(A,B)$ makes conditions comparable.
Finding: H2 is supported. Assimilation strength toward the majority decreases monotonically as the minority grows, on both the behavioral ($\Delta d_d$) and semantic ($\Delta d_o$) dimensions, with large effect sizes. The continuous metric adds what the bins could not: the change from $\alpha=0.12$ to $0.25$ is not just fewer assimilation runs but weaker assimilation in each. At $\alpha=0.50$ the distribution is centred on zero, confirming the slight "Population 2 advantage" seen in the table was stochastic drift, not a real tendency. Notably the semantic effect is the stronger of the two: because ontologies only grow, every agent keeps its origin's structural fingerprint, anchoring the merged ontology toward whichever population is more numerous.
Two claims are tested: lower connectivity should (H3.1) reduce majority assimilation / increase hybridization, and (H3.2) leave a more semantically diverse final culture.
Since H2 showed population size dominates, connectivity is first isolated in the balanced case ($\alpha = 0.5$), where the size effect is neutralised.
In the balanced case there is no effect of connectivity. The hypothesis expects a stronger effect in the unbalanced configurations, examined next.
The heatmaps below give the clearest view: mean distance difference across every combination of $\alpha$ (rows) and $\eta$ (columns), behaviorally and then semantically.
The interesting metric here is the ontology distance. Behavioral convergence is already guaranteed by H1 (everyone ends up deciding like the majority), so a behavioral measure can only restate that. The semantic measure can in principle reveal whether restricted connectivity let the minority's structure survive even after its decisions aligned.
Finding: H3.1 is not supported. The heatmaps show clean horizontal bands: colour barely changes along a row (varying $\eta$) but shifts sharply between rows (varying $\alpha$). Population size drives the outcome; connectivity does not, in either dimension, at any size. A connectivity signal hinted at in an earlier 5-run pilot did not survive the larger 10-run design. The reason is temporal: a sparse network may slow convergence, but given a long-enough budget the pressure to agree (H1) overrides any buffering, and because ontologies only grow the minority ends up accumulating the same adaptations regardless of topology.
Here "diversity" is the internal heterogeneity of the merged population: the average pairwise ontology distance among all its agents at the end of the run.
Finding: H3.2 is not supported. Mean final diversity is essentially flat across connectivity levels (≈ 0.476–0.489) and the difference is not significant. This null result runs counter to the single-population literature, where network fragmentation preserves diversity; here the hard constraint to reach agreement appears strong enough to wash that effect out. It may also signal that the decision task is simply not complex enough for connectivity to matter — a candidate for future work.
Two claims: higher overlap should (H4.1) shift outcomes toward hybridization, and (H4.2) lower the final semantic diversity. Both are analysed in the balanced case ($\alpha = 0.5$).
There is a visible jump in hybridization at full overlap ($\omega = 1$); the extreme conditions (0.0 vs 1.0) are compared directly to test it.
The continuous metric clarifies what is happening at $\omega = 1$.
The behavioral band explodes at high overlap while the semantic band stays narrow.
This is a denominator artifact: with full overlap the two populations are already almost identical before they meet, so the pre-encounter distance $d_d(h_1,h_2)$ shrinks toward zero and the normalised metric amplifies any tiny drift. The raw (non-normalised) difference confirms this, its spread shrinks as overlap grows, the exact opposite of the normalised version.
Finding: H4.1 is not supported. Neither the geometric scheme nor the raw distance difference shows any genuine, directional effect of overlap on the outcome at balanced sizes. The apparent spike at $\omega = 1$ is an artifact of the near-zero baseline distance — and conceptually, when two populations are already identical, "hybridization" loses its meaning: the encounter only settles a few leftover disagreements.
Finding: H4.2 is not supported. Higher overlap did not lower final diversity.
Instead, it was observed that the completely disjoint condition ($\omega = 0$) produced the least diverse final culture (mean ≈ 0.456 vs ≈ 0.49 for all overlapping conditions), a threshold effect separating only the disjoint case from the rest, significant with a small effect size.
The hypothesis assumed: more overlap → more similar before contact → fewer adaptations → less diverse result. The first link does hold — let's verify it before interpreting.
Interpretation. The pre-encounter distance does fall monotonically with overlap (dashed line): more shared training really does make populations more alike before contact. But final diversity (solid line) is lowest exactly where pre-encounter distance was highest, the paradox.
The resolution is in what kind of adaptation occurs, not how much. With disjoint training ($\omega = 0$), the two populations decide differently on a large share of objects, so the encounter forces a deep, system-wide renegotiation: agents rebuild compatible knowledge almost from scratch and converge onto a uniform shared structure, uniform, hence less internally diverse.
With any overlap, they already agree on much, so the encounter only needs targeted touch-ups, and each population largely keeps the structural diversity it built during stabilization.
However, this finding is only marginally significant and should be treated as an observation requiring further investigation to be conclusively verified.
| Hypothesis | Prediction | Outcome |
|---|---|---|
| H1 | Encounter → behavioral consensus | Supported |
| H2 | Larger population assimilates smaller | Supported |
| H3.1 | Lower connectivity → less assimilation | Not supported |
| H3.2 | Lower connectivity → more diversity | Not supported |
| H4.1 | Higher overlap → more hybridization | Not supported |
| H4.2 | Higher overlap → less diversity | Not supported |
Bottom line. The outcome of a cultural encounter in this model is governed almost entirely by relative population size: the larger group assimilates the smaller, more strongly the more lopsided the sizes. Neither the structure of contact nor the amount of shared prior knowledge had a comparable effect, contrary to what single-population studies would suggest.
A reproducibility safeguard: the simulator version (Lazy Lavender hash) recorded with the run is compared against the version requested in the parameter file.