On this tutorial, we construct a safety-critical reinforcement studying pipeline that learns fully from fastened, offline knowledge slightly than dwell exploration. We design a customized atmosphere, generate a conduct dataset from a constrained coverage, after which practice each a Habits Cloning baseline and a Conservative Q-Studying agent utilizing d3rlpy. By structuring the workflow round offline datasets, cautious analysis, and conservative studying goals, we display how sturdy decision-making insurance policies could be educated in settings the place unsafe exploration isn’t an possibility. Take a look at the FULL CODES here.
!pip -q set up -U "d3rlpy" "gymnasium" "numpy" "torch" "matplotlib" "scikit-learn"
import os
import time
import random
import examine
import numpy as np
import matplotlib.pyplot as plt
import gymnasium as health club
from gymnasium import areas
import torch
import d3rlpy
SEED = 42
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
def pick_device():
if torch.cuda.is_available():
return "cuda:0"
return "cpu"
DEVICE = pick_device()
print("d3rlpy:", getattr(d3rlpy, "__version__", "unknown"), "| torch:", torch.__version__, "| system:", DEVICE)
def make_config(cls, **kwargs):
sig = examine.signature(cls.__init__)
allowed = set(sig.parameters.keys())
allowed.discard("self")
filtered = {ok: v for ok, v in kwargs.gadgets() if ok in allowed}
return cls(**filtered)We arrange the atmosphere by putting in dependencies, importing libraries, and fixing random seeds for reproducibility. We detect and configure the computation system to make sure constant execution throughout methods. We additionally outline a utility to create configuration objects safely throughout completely different d3rlpy variations. Take a look at the FULL CODES here.
class SafetyCriticalGridWorld(health club.Env):
metadata = {"render_modes": []}
def __init__(
self,
dimension=15,
max_steps=80,
hazard_coords=None,
begin=(0, 0),
objective=None,
slip_prob=0.05,
seed=0,
):
tremendous().__init__()
self.dimension = int(dimension)
self.max_steps = int(max_steps)
self.begin = tuple(begin)
self.objective = tuple(objective) if objective isn't None else (self.dimension - 1, self.dimension - 1)
self.slip_prob = float(slip_prob)
if hazard_coords is None:
hz = set()
rng = np.random.default_rng(seed)
for _ in vary(max(1, self.dimension // 2)):
x = rng.integers(2, self.dimension - 2)
y = rng.integers(2, self.dimension - 2)
hz.add((int(x), int(y)))
self.hazards = hz
else:
self.hazards = set(tuple(x) for x in hazard_coords)
self.action_space = areas.Discrete(4)
self.observation_space = areas.Field(low=0.0, excessive=float(self.dimension - 1), form=(2,), dtype=np.float32)
self._rng = np.random.default_rng(seed)
self._pos = None
self._t = 0
def reset(self, *, seed=None, choices=None):
if seed isn't None:
self._rng = np.random.default_rng(seed)
self._pos = [int(self.start[0]), int(self.begin[1])]
self._t = 0
obs = np.array(self._pos, dtype=np.float32)
return obs, {}
def _clip(self):
self._pos[0] = int(np.clip(self._pos[0], 0, self.dimension - 1))
self._pos[1] = int(np.clip(self._pos[1], 0, self.dimension - 1))
def step(self, motion):
self._t += 1
a = int(motion)
if self._rng.random() < self.slip_prob:
a = int(self._rng.integers(0, 4))
if a == 0:
self._pos[1] += 1
elif a == 1:
self._pos[0] += 1
elif a == 2:
self._pos[1] -= 1
elif a == 3:
self._pos[0] -= 1
self._clip()
x, y = int(self._pos[0]), int(self._pos[1])
terminated = False
truncated = self._t >= self.max_steps
reward = -1.0
if (x, y) in self.hazards:
reward = -100.0
terminated = True
if (x, y) == self.objective:
reward = +50.0
terminated = True
obs = np.array([x, y], dtype=np.float32)
return obs, float(reward), terminated, truncated, {}We outline a safety-critical GridWorld atmosphere with hazards, terminal states, and stochastic transitions. We encode penalties for unsafe states and rewards for profitable activity completion. We make sure the atmosphere strictly controls dynamics to mirror real-world security constraints. Take a look at the FULL CODES here.
def safe_behavior_policy(obs, env: SafetyCriticalGridWorld, epsilon=0.15):
x, y = int(obs[0]), int(obs[1])
gx, gy = env.objective
most well-liked = []
if gx > x:
most well-liked.append(1)
elif gx < x:
most well-liked.append(3)
if gy > y:
most well-liked.append(0)
elif gy < y:
most well-liked.append(2)
if len(most well-liked) == 0:
most well-liked = [int(env._rng.integers(0, 4))]
if env._rng.random() < epsilon:
return int(env._rng.integers(0, 4))
candidates = []
for a in most well-liked:
nx, ny = x, y
if a == 0:
ny += 1
elif a == 1:
nx += 1
elif a == 2:
ny -= 1
elif a == 3:
nx -= 1
nx = int(np.clip(nx, 0, env.dimension - 1))
ny = int(np.clip(ny, 0, env.dimension - 1))
if (nx, ny) not in env.hazards:
candidates.append(a)
if len(candidates) == 0:
return most well-liked[0]
return int(random.alternative(candidates))
def generate_offline_episodes(env, n_episodes=400, epsilon=0.20, seed=0):
episodes = []
for i in vary(n_episodes):
obs, _ = env.reset(seed=int(seed + i))
obs_list = []
act_list = []
rew_list = []
done_list = []
accomplished = False
whereas not accomplished:
a = safe_behavior_policy(obs, env, epsilon=epsilon)
nxt, r, terminated, truncated, _ = env.step(a)
accomplished = bool(terminated or truncated)
obs_list.append(np.array(obs, dtype=np.float32))
act_list.append(np.array([a], dtype=np.int64))
rew_list.append(np.array([r], dtype=np.float32))
done_list.append(np.array([1.0 if done else 0.0], dtype=np.float32))
obs = nxt
episodes.append(
{
"observations": np.stack(obs_list, axis=0),
"actions": np.stack(act_list, axis=0),
"rewards": np.stack(rew_list, axis=0),
"terminals": np.stack(done_list, axis=0),
}
)
return episodes
def build_mdpdataset(episodes):
obs = np.concatenate([ep["observations"] for ep in episodes], axis=0).astype(np.float32)
acts = np.concatenate([ep["actions"] for ep in episodes], axis=0).astype(np.int64)
rews = np.concatenate([ep["rewards"] for ep in episodes], axis=0).astype(np.float32)
phrases = np.concatenate([ep["terminals"] for ep in episodes], axis=0).astype(np.float32)
if hasattr(d3rlpy, "dataset") and hasattr(d3rlpy.dataset, "MDPDataset"):
return d3rlpy.dataset.MDPDataset(observations=obs, actions=acts, rewards=rews, terminals=phrases)
increase RuntimeError("d3rlpy.dataset.MDPDataset not discovered. Improve d3rlpy.")We design a constrained conduct coverage that generates offline knowledge with out dangerous exploration. We roll out this coverage to gather trajectories and construction them into episodes. We then convert these episodes right into a format appropriate with d3rlpy’s offline studying APIs. Take a look at the FULL CODES here.
def _get_episodes_from_dataset(dataset):
if hasattr(dataset, "episodes") and dataset.episodes isn't None:
return dataset.episodes
if hasattr(dataset, "get_episodes"):
return dataset.get_episodes()
increase AttributeError("Couldn't discover episodes in dataset (d3rlpy model mismatch).")
def _iter_all_observations(dataset):
for ep in _get_episodes_from_dataset(dataset):
obs = getattr(ep, "observations", None)
if obs is None:
proceed
for o in obs:
yield o
def _iter_all_transitions(dataset):
for ep in _get_episodes_from_dataset(dataset):
obs = getattr(ep, "observations", None)
acts = getattr(ep, "actions", None)
rews = getattr(ep, "rewards", None)
if obs is None or acts is None:
proceed
n = min(len(obs), len(acts))
for i in vary(n):
o = obs[i]
a = acts[i]
r = rews[i] if rews isn't None and that i < len(rews) else None
yield o, a, r
def visualize_dataset(dataset, env, title="Offline Dataset"):
state_visits = np.zeros((env.dimension, env.dimension), dtype=np.float32)
for obs in _iter_all_observations(dataset):
x, y = int(obs[0]), int(obs[1])
x = int(np.clip(x, 0, env.dimension - 1))
y = int(np.clip(y, 0, env.dimension - 1))
state_visits[y, x] += 1
plt.determine(figsize=(6, 5))
plt.imshow(state_visits, origin="decrease")
plt.colorbar(label="Visits")
plt.scatter([env.start[0]], [env.start[1]], marker="o", label="begin")
plt.scatter([env.goal[0]], [env.goal[1]], marker="*", label="objective")
if len(env.hazards) > 0:
hz = np.array(listing(env.hazards), dtype=np.int32)
plt.scatter(hz[:, 0], hz[:, 1], marker="x", label="hazards")
plt.title(f"{title} — State visitation")
plt.xlabel("x")
plt.ylabel("y")
plt.legend()
plt.present()
rewards = []
for _, _, r in _iter_all_transitions(dataset):
if r isn't None:
rewards.append(float(r))
if len(rewards) > 0:
plt.determine(figsize=(6, 4))
plt.hist(rewards, bins=60)
plt.title(f"{title} — Reward distribution")
plt.xlabel("reward")
plt.ylabel("depend")
plt.present()We implement dataset utilities that accurately iterate by means of episodes slightly than assuming flat arrays. We visualize state visitation to know protection and knowledge bias within the offline dataset. We additionally analyze reward distributions to examine the training sign accessible to the agent. Take a look at the FULL CODES here.
def rollout_eval(env, algo, n_episodes=25, seed=0):
returns = []
lengths = []
hazard_hits = 0
goal_hits = 0
for i in vary(n_episodes):
obs, _ = env.reset(seed=seed + i)
accomplished = False
complete = 0.0
steps = 0
whereas not accomplished:
a = int(algo.predict(np.asarray(obs, dtype=np.float32)[None, ...])[0])
obs, r, terminated, truncated, _ = env.step(a)
complete += float(r)
steps += 1
accomplished = bool(terminated or truncated)
if terminated:
x, y = int(obs[0]), int(obs[1])
if (x, y) in env.hazards:
hazard_hits += 1
if (x, y) == env.objective:
goal_hits += 1
returns.append(complete)
lengths.append(steps)
return {
"return_mean": float(np.imply(returns)),
"return_std": float(np.std(returns)),
"len_mean": float(np.imply(lengths)),
"hazard_rate": float(hazard_hits / max(1, n_episodes)),
"goal_rate": float(goal_hits / max(1, n_episodes)),
"returns": np.asarray(returns, dtype=np.float32),
}
def action_mismatch_rate_vs_data(dataset, algo, sample_obs=7000, seed=0):
rng = np.random.default_rng(seed)
obs_all = []
act_all = []
for o, a, _ in _iter_all_transitions(dataset):
obs_all.append(np.asarray(o, dtype=np.float32))
act_all.append(int(np.asarray(a).reshape(-1)[0]))
if len(obs_all) >= 80_000:
break
obs_all = np.stack(obs_all, axis=0)
act_all = np.asarray(act_all, dtype=np.int64)
idx = rng.alternative(len(obs_all), dimension=min(sample_obs, len(obs_all)), exchange=False)
obs_probe = obs_all[idx]
act_probe_data = act_all[idx]
act_probe_pi = algo.predict(obs_probe).astype(np.int64)
mismatch = (act_probe_pi != act_probe_data).astype(np.float32)
return float(mismatch.imply())
def create_discrete_bc(system):
if hasattr(d3rlpy.algos, "DiscreteBCConfig"):
cls = d3rlpy.algos.DiscreteBCConfig
cfg = make_config(
cls,
learning_rate=3e-4,
batch_size=256,
)
return cfg.create(system=system)
if hasattr(d3rlpy.algos, "DiscreteBC"):
return d3rlpy.algos.DiscreteBC()
increase RuntimeError("DiscreteBC not accessible on this d3rlpy model.")
def create_discrete_cql(system, conservative_weight=6.0):
if hasattr(d3rlpy.algos, "DiscreteCQLConfig"):
cls = d3rlpy.algos.DiscreteCQLConfig
cfg = make_config(
cls,
learning_rate=3e-4,
actor_learning_rate=3e-4,
critic_learning_rate=3e-4,
temp_learning_rate=3e-4,
alpha_learning_rate=3e-4,
batch_size=256,
conservative_weight=float(conservative_weight),
n_action_samples=10,
rollout_interval=0,
)
return cfg.create(system=system)
if hasattr(d3rlpy.algos, "DiscreteCQL"):
algo = d3rlpy.algos.DiscreteCQL()
if hasattr(algo, "conservative_weight"):
strive:
algo.conservative_weight = float(conservative_weight)
besides Exception:
move
return algo
increase RuntimeError("DiscreteCQL not accessible on this d3rlpy model.")We outline managed analysis routines to measure coverage efficiency with out uncontrolled exploration. We compute returns and security metrics, together with hazard and objective charges. We additionally introduce a mismatch diagnostic to quantify how usually realized actions deviate from the dataset conduct. Take a look at the FULL CODES here.
def most important():
env = SafetyCriticalGridWorld(
dimension=15,
max_steps=80,
slip_prob=0.05,
seed=SEED,
)
raw_eps = generate_offline_episodes(env, n_episodes=500, epsilon=0.22, seed=SEED)
dataset = build_mdpdataset(raw_eps)
print("dataset constructed:", kind(dataset).__name__)
visualize_dataset(dataset, env, title="Habits Dataset (Offline)")
bc = create_discrete_bc(DEVICE)
cql = create_discrete_cql(DEVICE, conservative_weight=6.0)
print("nTraining Discrete BC (offline)...")
t0 = time.time()
bc.match(
dataset,
n_steps=25_000,
n_steps_per_epoch=2_500,
experiment_name="grid_bc_offline",
)
print("BC practice sec:", spherical(time.time() - t0, 2))
print("nTraining Discrete CQL (offline)...")
t0 = time.time()
cql.match(
dataset,
n_steps=80_000,
n_steps_per_epoch=8_000,
experiment_name="grid_cql_offline",
)
print("CQL practice sec:", spherical(time.time() - t0, 2))
print("nControlled on-line analysis (small variety of rollouts):")
bc_metrics = rollout_eval(env, bc, n_episodes=30, seed=SEED + 1000)
cql_metrics = rollout_eval(env, cql, n_episodes=30, seed=SEED + 2000)
print("BC :", {ok: v for ok, v in bc_metrics.gadgets() if ok != "returns"})
print("CQL:", {ok: v for ok, v in cql_metrics.gadgets() if ok != "returns"})
print("nOOD-ish diagnostic (coverage motion mismatch vs knowledge motion at similar states):")
bc_mismatch = action_mismatch_rate_vs_data(dataset, bc, sample_obs=7000, seed=SEED + 1)
cql_mismatch = action_mismatch_rate_vs_data(dataset, cql, sample_obs=7000, seed=SEED + 2)
print("BC mismatch charge :", bc_mismatch)
print("CQL mismatch charge:", cql_mismatch)
plt.determine(figsize=(6, 4))
labels = ["BC", "CQL"]
means = [bc_metrics["return_mean"], cql_metrics["return_mean"]]
stds = [bc_metrics["return_std"], cql_metrics["return_std"]]
plt.bar(labels, means, yerr=stds)
plt.ylabel("Return")
plt.title("On-line Rollout Return (Managed)")
plt.present()
plt.determine(figsize=(6, 4))
plt.plot(np.kind(bc_metrics["returns"]), label="BC")
plt.plot(np.kind(cql_metrics["returns"]), label="CQL")
plt.xlabel("Episode (sorted)")
plt.ylabel("Return")
plt.title("Return Distribution (Sorted)")
plt.legend()
plt.present()
out_dir = "/content material/offline_rl_artifacts"
os.makedirs(out_dir, exist_ok=True)
bc_path = os.path.be part of(out_dir, "grid_bc_policy.pt")
cql_path = os.path.be part of(out_dir, "grid_cql_policy.pt")
if hasattr(bc, "save_policy"):
bc.save_policy(bc_path)
print("Saved BC coverage:", bc_path)
if hasattr(cql, "save_policy"):
cql.save_policy(cql_path)
print("Saved CQL coverage:", cql_path)
print("nDone.")
if __name__ == "__main__":
most important()We practice each Habits Cloning and Conservative Q-Studying brokers purely from offline knowledge. We examine their efficiency utilizing managed rollouts and diagnostic metrics. We finalize the workflow by saving educated insurance policies and summarizing safety-aware studying outcomes.
In conclusion, we demonstrated that Conservative Q-Studying yields a extra dependable coverage than easy imitation when studying from historic knowledge in safety-sensitive environments. By evaluating offline coaching outcomes, managed on-line evaluations, and action-distribution mismatches, we illustrated how conservatism helps scale back dangerous, out-of-distribution conduct. General, we introduced a whole, reproducible offline RL workflow that we are able to prolong to extra complicated domains comparable to robotics, healthcare, or finance with out compromising security.
Take a look at the FULL CODES here. Additionally, be happy to observe us on Twitter and don’t neglect to hitch our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
