Project DARWIN -- Diving Into Reinforcement Learning

Visit Individual Folders for Demo Videos!

• Project Inspired by OpenAI's "Emergent Tool Use from Multi-Agent Autocurricula" Link Here

RL in a Nutshell

Example Training Loop

env = gym.make(
        "LunarLander-v2",
        continuous = True,
        gravity = -10.0,
        render_mode = None
    )

agent = Agent(alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, env=env, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4)

np.random.seed(0)
score_history = []

for i in range(1000):
    done = False
    score = 0
    obs, _ = env.reset()
    while not done:
        print(obs.shape)
        act = agent.choose_action(obs)
        new_state, reward, terminated, truncated, info = env.step(act)
        done = terminated or truncated
        agent.remember(obs, act, reward, new_state, int(done))
        agent.learn()
        score += reward
        obs = new_state

    score_history.append(score)
    print("episode", i, "score %.2f" % score, "100 game average %.2f" % np.mean(score_history[-100:]))
    if i % 25 == 0:
        agent.save_models()

The Team