By Killian Trouillet

During training, we used GAMA in headless mode for speed. Now that the model is trained, we switch to the GAMA GUI to watch the agent in action.

1. Open GAMA normally (double-click the application)
2. The GUI server runs on port 1000 by default — no extra setup needed

from train_forager import PPOAgent

agent = PPOAgent(state_dim=13, action_dim=2)
agent.load("saved_models/ppo_forager.pth")

This loads the neural network weights saved during training.

The only change from training: the port number.

env = gym.make(
    "gama_gymnasium_env/GamaEnv-v0",
    gaml_experiment_path="path/to/forager_gym.gaml",
    gaml_experiment_name="gym_env",
    gama_ip_address="localhost",
    gama_port=1000,
)

action, _, _ = agent.select_action(obs, test=True)

During training, the policy samples actions from a Normal distribution (adding noise for exploration). During testing, we set test=True to use the mean action — no randomness, just the best action the network has learned.

time.sleep(0.1)  # 100ms delay between steps

We add a small delay between steps so the movement is visible in the GAMA display.

cd models/gym
python test_forager.py

Watch the GAMA display — the blue forager should navigate smoothly around obstacles to reach the green food!

==================================================
  Smart Forager - PPO Test (GUI)
==================================================
  Step 0: obs[:5]=[0.05, 0.05, 0.95, 0.52, 0.48] action=[0.73, 0.71]
  Step 1: obs[:5]=[0.07, 0.07, 0.93, 0.54, 0.50] action=[0.81, 0.78]
  ...

  Result: FOUND FOOD! | Steps: 49 | Reward: 92.6

Done.

global, species, reflex, action, aspect, experiment, parameter, geometry, rectangle, circle, line, distance_to, towards, intersects, inter, cos, sin, list, map, loop, draw

gymnasium, gym.make(), env.reset(), env.step(), torch.nn.Module, Normal distribution, ActorCritic, PPOAgent, asyncio