16. Stress Test — 10 Random Agents

By Killian Trouillet

---

Content

In Steps 12–15 we trained and tested a cooperative policy with 2 foragers. Because we used parameter sharing, the single trained network generalises to any number of agents — each one independently runs its own forward pass through the same weights.

In this last step, we evaluate that claim without any re-training: we create a new GAML model with 10 foragers placed across the map and run the trained model by flipping one flag in the test script.

---

What Changes

Aspect	Step 15 (Normal Test)	Step 16 (Stress Test)
Agents	2 (fixed positions)	10 (spread across the map)
GAML model	forager_petz.gaml	forager_petz_stress.gaml
Experiment	petz_env	petz_stress_env
Python script	test_forager_petz.py	same script — STRESS_TEST = True
Model	Shared PPO	Same shared PPO
max_steps	300	500 (extra time for 10 agents)

---

The New GAML Model

forager_petz_stress.gaml is a self-contained model derived from forager_petz.gaml with three differences.

10 Agent IDs

list<string> agent_ids <- [
    "forager_0", "forager_1", "forager_2", "forager_3", "forager_4",
    "forager_5", "forager_6", "forager_7", "forager_8", "forager_9"
];
list<bool> agents_at_food <- [
    false, false, false, false, false,
    false, false, false, false, false
];

10 Varied Starting Positions

list<point> start_positions <- [
    {5.0,  5.0},   // top-left corner
    {15.0, 5.0},
    {5.0, 15.0},
    {50.0, 5.0},   // top-center
    {5.0, 50.0},   // left-center
    {50.0, 50.0},  // center
    {5.0, 85.0},   // bottom-left
    {50.0, 85.0},  // bottom-center
    {85.0, 5.0},   // top-right
    {85.0, 50.0}   // right-center
];

Positions are chosen to be free of obstacles, spread across all map quadrants.

Longer Episode Limit

int max_steps <- 500;  // 300 in the 2-agent model

With 10 agents spread across the map, some start far from the food. 500 steps gives even distant agents a fair chance.

Observation: Nearest Teammate

The trained model still expects 15 values (2 for teammate position). With 10 agents, compute_observation reports the nearest other agent instead of the first one:

forager nearest <- (forager where (each.agent_id != agent_id)) closest_to self;
if (nearest != nil) {
    other_x <- nearest.location.x / world_size;
    other_y <- nearest.location.y / world_size;
}

This is a zero-shot adaptation — the trained weights are unchanged.

Aspect: 10 Distinct Colours

list<rgb> palette <- [
    #blue, rgb(0,180,180), rgb(0,180,0), rgb(180,0,180), rgb(180,100,0),
    rgb(0,100,200), rgb(200,0,100), rgb(100,200,0), rgb(80,80,200), rgb(200,80,80)
];
rgb agent_color <- palette[agent_index mod length(palette)];
draw circle(0.8) color: at_food ? #orange : agent_color;

Each forager gets a distinct colour. When it reaches the food it turns orange, just as in the 2-agent case.

---

Running the Stress Test

Step 1 – Open GAMA GUI

Open GAMA normally (port 1000).

Step 2 – Enable STRESS_TEST

Open test_forager_petz.py and change the one flag at the top:

STRESS_TEST = True   # ← was False

Step 3 – Run

cd models/petz
python test_forager_petz.py

Expected Console Output

=======================================================
  Smart Forager — MARL Test (gama-pettingzoo GUI)
=======================================================
Model loaded (shared by both foragers)
  Connecting to GAMA (attempt 1/8)...
  Connected!

Running 1 cooperative test episodes...

  Episode 1/1: ✓ COOPERATIVE SUCCESS! | Steps: 134
    forager_0: reward = 86.3
    forager_1: reward = 84.7
    forager_2: reward = 79.1
    forager_3: reward = 81.4
    forager_4: reward = 77.8
    forager_5: reward = 82.2
    forager_6: reward = 74.5
    forager_7: reward = 78.9
    forager_8: reward = 83.6
    forager_9: reward = 80.1

=======================================================
  Test Results Summary
=======================================================
  Episodes    : 1
  Success Rate: 100%
  Avg Steps   : 134
  forager_0 avg reward: 86.3
  ...
=======================================================

Note: with 10 agents all needing to reach the food, success requires the most distant agent to navigate the full map. Episodes that hit the 500-step timeout show a lower success rate — this is expected and does not mean the network failed to learn.

---

Key Files

File	Description
models/petz/forager_petz_stress.gaml	10-agent stress model
models/petz/test_forager_petz.py	Test script — set STRESS_TEST = True
models/petz/train_forager_petz.py	Training script (unchanged, 2 agents)