Deep Reinforcement Learning (Deep RL) is the machine learning paradigm where neural networks learn optimal behavior through trial-and-error interaction with an environment — receiving reward signals that guide policy improvement without labeled training data, enabling agents to master complex sequential decision-making tasks from game playing and robotics to resource allocation and chip design.

Core Framework

At each timestep t, an agent observes state s_t, takes action a_t according to policy π(a|s), receives reward r_t, and transitions to state s_{t+1}. The objective is to find the policy that maximizes cumulative discounted reward: E[Σ γ^t × r_t] where γ ∈ [0,1) is the discount factor.

Value-Based Methods

• DQN (Deep Q-Network): A CNN approximates the Q-function Q(s,a) — the expected cumulative reward of taking action a in state s. The agent acts greedily with respect to Q (choose action with highest Q-value). Experience replay (storing transitions in a buffer and sampling mini-batches) and target network (slowly updated copy of Q-network) stabilize training. Achieved superhuman Atari game play (DeepMind, 2015).
• Double DQN: Uses the online network to select the best action but the target network to evaluate it, reducing Q-value overestimation bias.
• Dueling DQN: Separates Q into state-value V(s) and advantage A(s,a) streams, improving learning when many actions have similar values.

Policy Gradient Methods

• REINFORCE: Directly parameterize the policy π_θ(a|s) and update θ by gradient ascent on expected reward. The policy gradient theorem: ∇J = E[∇log π_θ(a|s) × R_t]. Simple but high variance.
• PPO (Proximal Policy Optimization): Clips the policy ratio to prevent destructively large updates. The workhorse of modern deep RL — stable, sample-efficient, and easy to tune. Used for RLHF (RL from Human Feedback) in ChatGPT and other LLMs.
• Actor-Critic: The actor (policy network) selects actions; the critic (value network) estimates how good the current state is. The advantage (actual reward minus critic's estimate) reduces variance. A2C (synchronous), A3C (asynchronous multi-worker) scale to complex environments.

RLHF for Language Models

The application that brought deep RL to mainstream AI: 1. Supervised Fine-Tuning (SFT): Fine-tune the LLM on human-written demonstrations. 2. Reward Model Training: Train a reward model on human preference comparisons (response A vs. response B). 3. PPO Optimization: Use PPO to fine-tune the LLM to maximize the reward model's score while staying close to the SFT policy (KL penalty). Aligns the LLM with human preferences for helpfulness, harmlessness, and honesty.

Challenges

• Sample Efficiency: Deep RL typically requires millions of environment interactions. Sim-to-real transfer trains in simulation and deploys to the real world.
• Reward Specification: Designing reward functions that capture the true objective without unintended shortcuts (reward hacking) is notoriously difficult.
• Exploration: In sparse-reward environments, random exploration rarely discovers rewarding states. Intrinsic motivation (curiosity-driven exploration) and hierarchical RL address this.

Deep Reinforcement Learning is the framework for learning through interaction — the closest machine learning comes to how animals learn, discovering optimal strategies through experience rather than instruction, and now serving as the alignment mechanism that makes large language models useful and safe.