Question 56
How does Policy Iteration differ from Value Iteration in solving MDPs?
Policy Iteration explicitly evaluates the exact utility of policy $\pi$, then greedily improves $\pi , often converging$ in fewer iterations than Value Iteration.
How does Policy Iteration differ from Value Iteration in solving MDPs?
Policy Iteration explicitly evaluates the exact utility of policy $\pi$, then greedily improves $\pi , often converging$ in fewer iterations than Value Iteration.
Deep Q-Networks (DQN) stabilized deep neural network learning for Q-learning by introducing:
DQN uses Experience Replay to break temporal correlations in training samples and separate Target Networks to stabilize target Q-value updates.
In RL, an $\epsilon -greedy action selection strategy chooses a random exploratory action$ with probability:
In $\epsilon -greedy selection$, the agent selects a completely random exploratory action with probability $\epsilon$, and the current best greedy action with probability 1 - $\epsilon$.
Model-Free Reinforcement Learning algorithms are characterized by:
Model-free RL (e.g., Q-learning, Policy Gradients) learns from environmental experience samples without estimating environment transition dynamics.
What is the primary drawback of simple Hill Climbing search in complex state spaces?
Hill Climbing makes local greedy moves without backtracking, making it susceptible to getting trapped in local optima, flat plateaus, or narrow ridges.