"""
APPO agent for SUMO VSL with edge-structured tokenization.
"""
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from typing import Dict, List, Tuple


class SpatialAttentionBlock(nn.Module):
    """Self-attention block over ordered edge tokens."""

    def __init__(self, hidden_dim: int, num_heads: int = 4, dropout: float = 0.1):
        super().__init__()
        self.attn = nn.MultiheadAttention(
            embed_dim=hidden_dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
        )
        self.norm1 = nn.LayerNorm(hidden_dim)
        self.norm2 = nn.LayerNorm(hidden_dim)
        self.ffn = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim * 2),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim * 2, hidden_dim),
        )
        self.dropout = nn.Dropout(dropout)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        attn_out, _ = self.attn(x, x, x, need_weights=False)
        x = self.norm1(x + self.dropout(attn_out))
        ffn_out = self.ffn(x)
        return self.norm2(x + self.dropout(ffn_out))


class MultiDiscreteActorCritic(nn.Module):
    """Actor-critic that builds one token per controlled edge."""

    def __init__(
        self,
        state_dim: int,
        action_dims: List[int],
        edge_feature_dim: int = 3,
        time_feature_dim: int = 3,
        total_edge_count: int | None = None,
        controlled_start_index: int = 0,
        hidden_dim: int = 128,
        num_heads: int = 4,
        num_layers: int = 2,
        dropout: float = 0.1,
    ):
        super().__init__()
        self.state_dim = state_dim
        self.action_dims = action_dims
        self.num_zones = len(action_dims)
        self.edge_feature_dim = edge_feature_dim
        self.speed_feature_dim = 1
        self.time_feature_dim = time_feature_dim
        self.total_edge_count = int(total_edge_count if total_edge_count is not None else self.num_zones)
        self.controlled_start_index = int(controlled_start_index)
        self.controlled_end_index = self.controlled_start_index + self.num_zones
        if self.controlled_end_index > self.total_edge_count:
            raise ValueError("controlled action slice exceeds total edge count")
        self.last_reward_dim = 1
        self.global_feature_dim = self.time_feature_dim + self.last_reward_dim
        self.agent_id_dim = 1
        self.local_obs_dim = (
            self.edge_feature_dim
            + self.speed_feature_dim
            + self.global_feature_dim
            + self.agent_id_dim
        )

        self.local_encoder = nn.Sequential(
            nn.Linear(self.local_obs_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.GELU(),
        )
        self.pos_encoding = nn.Parameter(torch.zeros(1, self.num_zones, hidden_dim))
        self.attention_layers = nn.ModuleList(
            [SpatialAttentionBlock(hidden_dim, num_heads=num_heads, dropout=dropout) for _ in range(num_layers)]
        )

        self.actor_heads = nn.ModuleList([nn.Linear(hidden_dim, adim) for adim in action_dims])
        self.critic = nn.Sequential(
            nn.Linear(hidden_dim * 2 + self.global_feature_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, 1),
        )

        agent_ids = torch.linspace(0.0, 1.0, self.num_zones, dtype=torch.float32)
        self.register_buffer("agent_id_features", agent_ids.view(1, self.num_zones, 1))
        self._init_weights()

    def _init_weights(self):
        for module in self.modules():
            if isinstance(module, nn.Linear):
                nn.init.orthogonal_(module.weight, gain=np.sqrt(2))
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)

        for head in self.actor_heads:
            nn.init.orthogonal_(head.weight, gain=0.01)
        nn.init.orthogonal_(self.critic[-1].weight, gain=1.0)
        nn.init.normal_(self.pos_encoding, mean=0.0, std=0.02)

    def _build_local_tokens(self, state: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        if state.dim() == 1:
            state = state.unsqueeze(0)

        batch_size = state.size(0)
        edge_block = self.total_edge_count * self.edge_feature_dim
        speed_block_start = edge_block
        speed_block_end = speed_block_start + self.total_edge_count
        global_block_start = speed_block_end
        global_block_end = global_block_start + self.global_feature_dim

        edge_features = state[:, :edge_block].view(batch_size, self.total_edge_count, self.edge_feature_dim)
        edge_features = edge_features[:, self.controlled_start_index:self.controlled_end_index, :]
        local_speed_limits = state[:, speed_block_start:speed_block_end].view(batch_size, self.total_edge_count, 1)
        local_speed_limits = local_speed_limits[:, self.controlled_start_index:self.controlled_end_index, :]
        global_features = state[:, global_block_start:global_block_end]
        repeated_global = global_features.unsqueeze(1).expand(-1, self.num_zones, -1)
        agent_ids = self.agent_id_features.expand(batch_size, -1, -1)

        tokens = torch.cat([edge_features, local_speed_limits, repeated_global, agent_ids], dim=-1)
        return tokens, global_features

    def forward(self, state: torch.Tensor) -> Tuple[List[torch.Tensor], torch.Tensor]:
        tokens, global_features = self._build_local_tokens(state)
        x = self.local_encoder(tokens) + self.pos_encoding

        for attention_layer in self.attention_layers:
            x = attention_layer(x)

        logits_list = [head(x[:, idx, :]) for idx, head in enumerate(self.actor_heads)]
        pooled = torch.cat([x.mean(dim=1), x.max(dim=1).values, global_features], dim=-1)
        value = self.critic(pooled)
        return logits_list, value

    def get_value(self, state: torch.Tensor) -> torch.Tensor:
        _, value = self.forward(state)
        return value


class APPOAgent:
    """APPO agent for SUMO MultiDiscrete action space."""

    def __init__(
        self,
        state_dim: int,
        action_dims: List[int],
        edge_feature_dim: int = 3,
        time_feature_dim: int = 3,
        total_edge_count: int | None = None,
        controlled_start_index: int = 0,
        hidden_dim: int = 128,
        num_heads: int = 4,
        num_layers: int = 2,
        learning_rate: float = 3e-4,
        gamma: float = 0.99,
        gae_lambda: float = 0.95,
        clip_epsilon: float = 0.2,
        value_coef: float = 0.5,
        entropy_coef: float = 0.02,
        max_grad_norm: float = 0.5,
        ppo_epochs: int = 10,
        minibatch_size: int = 64,
        device: str = "cuda",
        lr_schedule: str = "cosine",
        total_episodes: int = 300,
    ):
        self.device = torch.device(device if torch.cuda.is_available() else "cpu")
        self.gamma = gamma
        self.gae_lambda = gae_lambda
        self.clip_epsilon = clip_epsilon
        self.value_coef = value_coef
        self.entropy_coef = entropy_coef
        self.max_grad_norm = max_grad_norm
        self.ppo_epochs = ppo_epochs
        self.minibatch_size = minibatch_size
        self.action_dims = action_dims

        self.policy = MultiDiscreteActorCritic(
            state_dim=state_dim,
            action_dims=action_dims,
            edge_feature_dim=edge_feature_dim,
            time_feature_dim=time_feature_dim,
            total_edge_count=total_edge_count,
            controlled_start_index=controlled_start_index,
            hidden_dim=hidden_dim,
            num_heads=num_heads,
            num_layers=num_layers,
        ).to(self.device)

        self.optimizer = optim.Adam(self.policy.parameters(), lr=learning_rate, eps=1e-5)
        if lr_schedule == "cosine":
            self.scheduler = optim.lr_scheduler.CosineAnnealingLR(
                self.optimizer,
                T_max=total_episodes,
                eta_min=learning_rate * 0.1,
            )
        else:
            self.scheduler = None

        self.reset_buffers()

    def reset_buffers(self):
        self.states = []
        self.actions = []
        self.rewards = []
        self.values = []
        self.log_probs = []
        self.dones = []

    def select_action(self, state: np.ndarray, deterministic: bool = False) -> Tuple[np.ndarray, float, float]:
        state_tensor = torch.FloatTensor(state).unsqueeze(0).to(self.device)

        with torch.no_grad():
            logits_list, value = self.policy(state_tensor)

        actions = []
        log_prob_total = 0.0

        for logits in logits_list:
            dist = torch.distributions.Categorical(logits=logits)
            if deterministic:
                action = torch.argmax(logits, dim=-1).item()
            else:
                action = dist.sample().item()
            actions.append(action)
            log_prob_total += dist.log_prob(torch.tensor(action, device=self.device)).item()

        return np.array(actions, dtype=np.int64), log_prob_total, value.item()

    def store_transition(self, state, action, reward, value, log_prob, done):
        self.states.append(state)
        self.actions.append(action)
        self.rewards.append(reward)
        self.values.append(value)
        self.log_probs.append(log_prob)
        self.dones.append(done)

    def compute_gae(self, next_value: float) -> Tuple[np.ndarray, np.ndarray]:
        advantages = []
        gae = 0.0

        for t in reversed(range(len(self.rewards))):
            next_val = next_value if t == len(self.rewards) - 1 else self.values[t + 1]
            delta = self.rewards[t] + self.gamma * next_val * (1 - self.dones[t]) - self.values[t]
            gae = delta + self.gamma * self.gae_lambda * (1 - self.dones[t]) * gae
            advantages.insert(0, gae)

        advantages = np.array(advantages, dtype=np.float32)
        returns = advantages + np.array(self.values, dtype=np.float32)
        return advantages, returns

    def update(self, next_value: float) -> Dict[str, float]:
        if len(self.states) == 0:
            return {}

        advantages, returns = self.compute_gae(next_value)
        advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-8)

        states = torch.FloatTensor(np.array(self.states)).to(self.device)
        actions = torch.LongTensor(np.array(self.actions)).to(self.device)
        old_log_probs = torch.FloatTensor(self.log_probs).to(self.device)
        advantages_t = torch.FloatTensor(advantages).to(self.device)
        returns_t = torch.FloatTensor(returns).to(self.device)

        total_policy_loss = 0.0
        total_value_loss = 0.0
        total_entropy = 0.0
        update_count = 0

        dataset_size = len(self.states)
        for _ in range(self.ppo_epochs):
            indices = np.random.permutation(dataset_size)

            for start_idx in range(0, dataset_size, self.minibatch_size):
                end_idx = min(start_idx + self.minibatch_size, dataset_size)
                batch_idx = indices[start_idx:end_idx]

                batch_states = states[batch_idx]
                batch_actions = actions[batch_idx]
                batch_old_lp = old_log_probs[batch_idx]
                batch_adv = advantages_t[batch_idx]
                batch_ret = returns_t[batch_idx]

                logits_list, values = self.policy(batch_states)

                new_log_probs = torch.zeros(len(batch_idx), device=self.device)
                entropy = torch.zeros(len(batch_idx), device=self.device)

                for i, logits in enumerate(logits_list):
                    dist = torch.distributions.Categorical(logits=logits)
                    new_log_probs += dist.log_prob(batch_actions[:, i])
                    entropy += dist.entropy()

                entropy_mean = entropy.mean()
                ratio = torch.exp(new_log_probs - batch_old_lp)
                surr1 = ratio * batch_adv
                surr2 = torch.clamp(ratio, 1 - self.clip_epsilon, 1 + self.clip_epsilon) * batch_adv
                policy_loss = -torch.min(surr1, surr2).mean()
                value_loss = nn.functional.mse_loss(values.squeeze(-1), batch_ret)
                loss = policy_loss + self.value_coef * value_loss - self.entropy_coef * entropy_mean

                self.optimizer.zero_grad()
                loss.backward()
                nn.utils.clip_grad_norm_(self.policy.parameters(), self.max_grad_norm)
                self.optimizer.step()

                total_policy_loss += policy_loss.item()
                total_value_loss += value_loss.item()
                total_entropy += entropy_mean.item()
                update_count += 1

        if self.scheduler is not None:
            self.scheduler.step()
        self.reset_buffers()

        return {
            "policy_loss": total_policy_loss / max(update_count, 1),
            "value_loss": total_value_loss / max(update_count, 1),
            "entropy": total_entropy / max(update_count, 1),
        }

    def save(self, path: str):
        torch.save(
            {
                "policy_state_dict": self.policy.state_dict(),
                "optimizer_state_dict": self.optimizer.state_dict(),
            },
            path,
        )

    def load(self, path: str):
        checkpoint = torch.load(path, map_location=self.device, weights_only=False)
        self.policy.load_state_dict(checkpoint["policy_state_dict"])
        self.optimizer.load_state_dict(checkpoint["optimizer_state_dict"])