Bài 8: SAC & Advanced Algorithms — Off-Policy Deep RL

Reinforcement Learning: Từ Cơ bản đến Nâng cao

Phần 2: Deep Reinforcement Learning — Neural Networks meet RL

xdev.asia

Giới thiệu

SAC (Soft Actor-Critic) là state-of-the-art off-policy algorithm cho continuous action spaces — kết hợp maximum entropy objective với actor-critic. SAC tự động tune temperature, ổn định và sample-efficient.

---

1. Maximum Entropy RL

$$\pi^* = \arg\max_\pi \sum_t \mathbb{E}_{\pi}[r(s_t, a_t) + \alpha \mathcal{H}(\pi(\cdot|s_t))]$$

Key insight: Balance reward maximization VÀ entropy (exploration) → robust policies.

---

2. SAC Components

class SACActorContinuous(nn.Module):
    """Gaussian policy for continuous actions"""
    def __init__(self, state_dim, action_dim, hidden=256):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(state_dim, hidden), nn.ReLU(),
            nn.Linear(hidden, hidden), nn.ReLU(),
        )
        self.mean = nn.Linear(hidden, action_dim)
        self.log_std = nn.Linear(hidden, action_dim)
    
    def forward(self, state):
        feat = self.net(state)
        mean = self.mean(feat)
        log_std = self.log_std(feat).clamp(-20, 2)
        return mean, log_std
    
    def sample(self, state):
        mean, log_std = self.forward(state)
        std = log_std.exp()
        dist = torch.distributions.Normal(mean, std)
        x = dist.rsample()  # Reparameterization trick
        action = torch.tanh(x)
        log_prob = dist.log_prob(x) - torch.log(1 - action.pow(2) + 1e-6)
        return action, log_prob.sum(-1)

Twin Q-Networks

class TwinQNetwork(nn.Module):
    def __init__(self, state_dim, action_dim, hidden=256):
        super().__init__()
        self.q1 = nn.Sequential(
            nn.Linear(state_dim + action_dim, hidden), nn.ReLU(),
            nn.Linear(hidden, hidden), nn.ReLU(),
            nn.Linear(hidden, 1),
        )
        self.q2 = nn.Sequential(
            nn.Linear(state_dim + action_dim, hidden), nn.ReLU(),
            nn.Linear(hidden, hidden), nn.ReLU(),
            nn.Linear(hidden, 1),
        )
    
    def forward(self, state, action):
        x = torch.cat([state, action], dim=-1)
        return self.q1(x), self.q2(x)

---

3. TD3 — Twin Delayed DDPG

3 tricks trên DDPG:

1. Twin critics: Take minimum → reduce overestimation
2. Delayed policy updates: Update actor ít hơn critic
3. Target policy smoothing: Add noise to target actions

---

4. Algorithm Selection Guide

Algorithm	Action Space	On/Off-Policy	Sample Efficiency	Best For
DQN	Discrete	Off-policy	Good	Games
PPO	Both	On-policy	Low	General purpose
SAC	Continuous	Off-policy	High	Robotics
TD3	Continuous	Off-policy	High	Robotics
DDPG	Continuous	Off-policy	Good	Simple continuous

---

5. Model-based RL Overview

Thay vì model-free, học model of environment → plan:

Approach	Learn Model?	Plan?	Example
Model-free	No	No	DQN, PPO
Model-based	Yes	Yes	Dreamer, MuZero
Hybrid	Yes	Yes	World Models

---

Tổng kết

Concept	Mô tả
SAC	Max entropy + twin critics + auto temperature
TD3	Twin critics + delayed + smoothing
Entropy	Encourages exploration, robust policies
Off-policy	Sample efficient, replay buffer