sim02 — Dynamic Fitness Landscape

Open the interactive visualization — renders the grid heatmap (play/pause, step slider) and all metric charts from results.json.

What It Tests

Whether agents that modify their fitness landscape (stigmergic niche construction) produce qualitatively different dynamics from agents on a static landscape.

Tests H4 (Dynamic Environment) and H8 (Computational Complexity Enables Open-Endedness).

Design

• Population of 200 agents on a 20x20 grid
• Each agent has a binary strategy string of length N=8
• Fitness depends on strategy, position, and landscape state
• NK-like fitness function with K=3 epistatic interactions per gene
• Two conditions:
  • Static: Fixed landscape. Agents adapt TO it.
  • Dynamic: Agents deposit stigmergic traces that modify fitness contributions. Traces decay at rate 0.005.

Key Results

Rerun 2026-07-27 after a fix to the trace term. The earlier numbers came from a trace bonus that was (a) added regardless of the agent's own strategy[i], so it was identical for every strategy at a cell and could not change which strategy won, and (b) unbounded, which drove dynamic mean fitness to 2488 against static's 0.77 — a 3224× scale difference that made the conditions incomparable. The bonus is now strategy-dependent and saturating (capped at TRACE_WEIGHT=0.5 per gene). See ../REVIEW.md §3.

Both conditions CONVERGE. The dynamic condition converges even harder.

Fitness ratio is now 1.44× (was 3224×), so the two conditions are on comparable scales and the comparison is meaningful. Both plateau by generation ~100 and neither moves for the remaining 4,900 generations (diversity range 0 in the final 1000 for both).

What happened in the dynamic condition:

1. Agents deposit traces on the gene channels they carry, raising fitness for those strategies at that location
2. Higher fitness → agents reproduce more → deposit more traces
3. Positive feedback concentrates the population; the trace field grows to ~27300
4. The population converges to 2 strategies (vs 4 in the static condition)
5. One trace cluster remains at the end
6. Fitness rises from 0.49 to 1.11 and then plateaus

Why this is instructive:

This is NOT a failure — it's a discovery. The simulation reveals that:

1. Stigmergy ALONE does NOT produce open-ended evolution. It can make convergence WORSE: the dynamic condition ends with half the diversity of the static one. The positive feedback in stigmergic traces narrows the population rather than opening it up. This is Heylighen's "groupthink / collective stupidity" criticism — the same amplification that exploits good solutions also amplifies bad ones. Note this conclusion survived the fix, but it now rests on a diversity difference (2 vs 4) rather than on a fitness number that was an artifact.

2. The trace→actor crossing (H7) is NOT automatic. Traces accumulate, form a cluster, and persist (0.74). But they don't become autonomous new-level actors — they remain fitness modifiers. The population develops a monoculture, not multi-scale structure.

3. Bounding the trace term matters as much as its decay rate. With an unbounded additive bonus the landscape term swamps the base landscape entirely (it contributed 99.97% of dynamic fitness before the fix) and the model stops being a fitness-landscape experiment at all. Any "dynamic landscape" term needs to stay commensurate with the base landscape it is supposed to be modifying.

4. What's missing for open-endedness: The simulation confirms that three additional mechanisms are needed:

   • Trace autonomy: Traces must develop their own dynamics, not just be passive fitness modifiers
   • Competing traces: Multiple trace types that compete/interact, not just one global trace field
   • Autopoietic crossing: A mechanism for accumulated traces to become self-maintaining structures with their own rules (the H7 phase transition)

Connection to Echo's Failure

Smith & Bedau (1997) found that Echo converges to simple trading ecologies. Our simulation shows the same convergence, even WITH stigmergic landscape modification. This confirms that adding stigmergy to a single-scale model doesn't produce multi-scale composition — it just changes the convergence dynamics.

The key insight: stigmergy is necessary but not sufficient (as argued in Session 3). The missing ingredient is the autopoietic crossing — the mechanism by which accumulated traces become self-maintaining new-level actors.

Next Steps

• Sim03: Add trace competition (multiple trace types that interact)
• Sim04: Add the autopoietic crossing mechanism (traces that develop self-maintenance)
• Sweep trace decay rate to find the optimal balance for structure formation
• Test whether traces can develop their own dynamics (not just modify fitness)

How to Run

cd ~/brain/artificial-life/simulations/sim02_dynamic_landscape
python3 sim02.py

Results are saved to results.json.