Skip to content

Commit 1d55409

Browse files
committed
documenting results
1 parent d413b72 commit 1d55409

1 file changed

Lines changed: 105 additions & 26 deletions

File tree

docs/RESULTS.md

Lines changed: 105 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -1,15 +1,11 @@
1-
# Experimental Results
1+
### Experimental Results
22

33
This document summarizes the key findings from the predator-prey cellular automaton experiments exploring the Hydra effect and self-organized criticality.
44

5-
## Phase 1: Critical Point Identification
6-
7-
### Objective
5+
#### Phase 1: Critical Point Identification
86

97
Locate the critical prey mortality rate where the population dynamics undergo a qualitative transition and test whether the system exhibits signatures of self-organized criticality at this point.
108

11-
### Methods
12-
139
We performed a parameter sweep over ```prey_death_rate``` $\in [0, 0.2]$ while holding other parameters fixed.
1410

1511
```
@@ -21,9 +17,7 @@ grid_size = 1000
2117
```
2218
Each simulation ran for 1000 warmup steps followed by 1000 measurement steps. The mean populations and cluster statiustics were recorded on the final grid states.
2319

24-
### Results
25-
26-
![Bifurcation Diagram](images/bifurcation.png)
20+
![Bifurcation Diagram](/Users/kimonanagnostopoulos/CSS_Project/images/bifurication.png)
2721

2822
The bifurication diagrams reveals a clear Hydra effect; increasing prey mortality leads to increasing prey population.
2923

@@ -42,7 +36,7 @@ The bifurication diagrams reveals a clear Hydra effect; increasing prey mortalit
4236

4337
The prey cluster size distributions were analyzed at the transition point to test for SOC. SOC theory predicts a power law distribution.
4438

45-
![Cluster Size Distribution](images/criticality_phase1.png)
39+
![Cluster Size Distribution](/Users/kimonanagnostopoulos/CSS_Project/images/criticality_phase1.png)
4640

4741
The log-log plot shows the probability distribution of prey cluster sizes at the critical point. Visual inspection suggests approximate power law scaling with $$\alpha \approx 2.29$$.
4842

@@ -65,19 +59,14 @@ where:
6559
Our results showed $\alpha = 2.29$ and $R = -1.3$. The negative $R$ value indicates that a lognormal distribution provides a better fit to the cluster size data than a power-law. This is evidence against true self-organized criticality. This might suggest that the system is near critical but not at a true critical point.
6660

6761

68-
## Phase 2: Evolutionary SOC Analsyis
62+
#### Phase 2: Evolutionary SOC Analsyis
6963

70-
### Objective
7164

7265
Test whether predator-prey system exhibits SOC by allowing prey mortality rates to evolve. Under the SOC hypothesis, evolution should drive the system toward the transition point identified in Phase 1.
7366

74-
### Methods
75-
7667
We enabled per-cell evolution of the prey death parameter. Each prey individual carries its own ```prey_death_rate``` value. The offsprings inherit their parents value with Gaussian mutation. Evolution occurs during the warmup steps, then frozen for the measurement.
7768

78-
### Results
79-
80-
![Evolution Convergence](images/phase2_evolution.png)
69+
![Evolution Convergence](/Users/kimonanagnostopoulos/CSS_Project/images/evo_plot.png)
8170

8271
**Right panel**: The evolved prey death rate as a function of initial prey death rate.
8372

@@ -93,18 +82,13 @@ The results showed unexpected convergence behavior:
9382

9483
The negative $R$ value is even more extreme than at the Phase 1 critical point ($R = -1.3$), indicating the evolved state is **further from criticality**, not closer.
9584

96-
9785
These results represent a significant puzzle. The system appears to self-organize. However, this convergence does not take place toward the transition point. At the moment, we do not have a definitive explanation for this behavior. A possible explanation could relate to evolution optimizing for individual fitness, not population level-properties like criticality.
9886

9987

100-
## Phase 3: Finite-size Scaling
101-
102-
### Objective:
88+
#### Phase 3: Finite-size Scaling
10389

10490
Investigate whethr the near-critical behavior observed in Phase 1 is a finite-size artifact. If the system is critical, scaling analysis across different system sizes should reveal universal behavior. If not, larger system should show stronger deviation from power-law scaling.
10591

106-
### Methods
107-
10892
We ran simulation at the critical point ($d_1 = 0.0955$) across six grid sizes $L \in \{50, 100, 250, 500, 1000, 2500\}$ with fixed parameters:
10993

11094
```
@@ -115,6 +99,8 @@ predator_death_rate = 0.05
11599

116100
For each system size, we aggregated all prey cluster sizes across replicates, fitted a lognormal distrbution using the maximum likelihood estimation, and computed power-law vs. lognormal likelihood ratio $R$.
117101

102+
---
103+
118104
#### NOTE: Lognormal Distribution
119105

120106
The lognormal distribution describes a variable whose logarith is normally distributed:
@@ -129,9 +115,102 @@ where:
129115

130116
The lognormal parameters were fitted using `scipy.stats.lognorm.fit()` with location fixed at zero.
131117

132-
### Results
118+
---
119+
120+
![Finite-Size Scaling](/Users/kimonanagnostopoulos/CSS_Project/images/fss_phase3.png)
121+
122+
The figure shows prey cluster size distributions at the critical point across all system sizes with lognormal fit parameters and likelihood ratios.
123+
124+
**Observations:**
125+
126+
**1. Lognormal parameters are stable across system sizes**
127+
128+
The fitted $\mu$ and $\sigma$ values converge to approximately $\mu \approx 1.36$ and $\sigma \approx 1.39$ for $L \geq 250$. This consistency indicates that the lognormal form is a robust description of the cluster size distribution, not a finite-size artifact.
129+
130+
131+
**2. Evidence against power-law grows with system size**
132+
133+
This is the crucial result: as $L$ increases, $R$ becomes *more negative*, not less:
134+
135+
$$R: -0.8 \rightarrow -6.9 \rightarrow -22.9 \rightarrow -46.2 \rightarrow -89.1 \rightarrow -224.4$$
136+
137+
If the system were truly critical with finite-size scaling corrections, we would expect $R \rightarrow 0$ as $L \rightarrow \infty$.
138+
139+
**3. Distributions do not collapse**
140+
141+
At a true critical point, rescaling cluster sizes by $L^{d_f}$ (where $d_f$ is the fractal dimension) should collapse all distributions onto a universal scaling function. The cluster size distribution is lognormal, which becomes more apparent with increasing system size. This result suggests that the underlying process involves multiplicative growth rather than the scale free branching processes that generate power laws.
142+
143+
144+
#### Phase 4 & 5: Sensitivity Analysis and Directed Hunting
145+
146+
We map the hydra effect across the full 4D parameter space to understand the conditions under which the Hydra effect occurs, the strength of the effect across different parameter regimes, and whether directed hunting alters the effect.
147+
148+
Phase 4 uses random neighbor selection while Phase 6 uses directed hunting where predators preferentially target prey neighbors.
149+
150+
We performed as sweep across the following model parameters:
151+
152+
```
153+
prey_birth
154+
prey_death
155+
predator_birth
156+
predator_death
157+
158+
```
159+
160+
For each parameter combination, we computed the equilibrium prey density curve and detected the Hydra effect by checking fro regions with a positive derivative:
161+
162+
$$\frac{d(\text{prey density})}{d(\text{prey death})} > 0$$
163+
164+
A curve exhibits the Hydra effect is prey density increases with mortality over any portion of the ```prey_death``` range.
165+
166+
To measure the stength of the Hydra effect, we fitted a qudratic function to the rising portion of each curve (truncated at the max derivative point):
167+
168+
$$\rho(d_1) = a \cdot d_1^2 + b \cdot d_1 + c$$
169+
170+
The quadratic coefficient $a$ captures how sharply the prey density rises and falls. Larger $|a|$ indicates a more pronounced Hydra effect with sharper peak. This apporach filters to only the Hydra effect portion, captures curvature, and provides a single scalar metric for comparison across parameter space.
171+
172+
### Results: Undirected Hunting (Phase 4)
173+
174+
![Hydra Curves Undirected](/Users/kimonanagnostopoulos/CSS_Project/images/hyrda_fitting_plot.png)
175+
176+
**Left panel**: All parameter combinations exhibiting the Hydra effect.
177+
**Right panel**: Curves truncated at the maximum derivative point, isolating the rising (Hydra) portion. Color indicates the maximum positive derivative—darker blue corresponds to steeper initial increases.
178+
179+
![3D Heatmap Undirected](/Users/kimonanagnostopoulos/CSS_Project/images/cube_phase4.png)
180+
181+
The 3D voxel plots show Hydra effect strength (quadratic coefficient) across the parameter space.
182+
183+
![Quadratic Coefficient Undirected](/Users/kimonanagnostopoulos/CSS_Project/images/coeffs_phase4.png)
184+
185+
**Left panel (Strong Hydra, coef > 15)**: Sharp, pronounced peaks with prey density increasing from ~0.3–0.6 to ~0.7–0.8 before collapsing.
186+
**Right panel (Moderate Hydra, coef 5–10)**: Gentler curves with broader peaks. The Hydra effect is present but less dramatic.
187+
188+
### Results: Directed Hunting (Phase 6)
189+
190+
![3D Heatmap Directed](/Users/kimonanagnostopoulos/CSS_Project/images/cube_phase6.png)
191+
192+
The same analysis with directed hunting enabled reveals a dramatically different pattern.
193+
194+
**Observations:**
195+
196+
1. **Much stronger Hydra effect**: Maximum quadratic coefficients reach ~40–50, roughly **double** the undirected case
197+
198+
2. **Shifted parameter region**: The strong-effect region extends to higher predator birth rates and shows a different structure
199+
200+
3. **More concentrated**: The effect is stronger where it occurs but may be more restricted in parameter space
201+
202+
![Quadratic Coefficient Directed](/Users/kimonanagnostopoulos/CSS_Project/images/coeffs_phase6.png)
203+
204+
**Left panel (Strong Hydra, coef > 15)**: Extremely sharp peaks with quadratic coefficients reaching 49–98. However, the curves are notably noisier than the undirected case, with irregular trajectories.
205+
**Right panel (Moderate Hydra, coef 5–10)**: More variable behavior compared to undirected hunting. Some curves show multiple peaks or irregular shapes.
206+
207+
The results indicate that directed hunting amplifies the Hydra effect but also introduces instability. Smarter predators paradoxically create conditions for stronger Hydra effects.
208+
209+
210+
211+
212+
213+
133214

134-
![Finite-Size Scaling](images/phase3_fss.png)
135215

136-
The figure shows
137216

0 commit comments

Comments
 (0)