In high-dimensional environments, which approach uses neural networks to estimate values instead of a tabular table?

In high-dimensional environments, which approach uses neural networks to estimate values instead of a tabular table?

• A. Temporal Difference (TD) method
• B. Q-Learning
• C. Monte Carlo method
• D. Deep Reinforcement Learning

Answer: (D)

When the environment is high-dimensional, you can’t keep a separate value for every state-action pair in a table. Instead, you use a function approximator that can generalize across many states. Neural networks serve this role well, allowing the agent to estimate value functions from complex inputs. Deep Reinforcement Learning uses these networks to approximate the value function (or Q-function) directly from raw observations, rather than storing values in a tabular form. A common example is a network that takes the current state and outputs a Q-value for each possible action, guiding decisions through learning signals like TD errors. This approach scales to rich inputs such as images or high-dimensional sensor data, where tabular methods would be impractical. While other methods involve temporal-difference updates or Monte Carlo returns, they don’t inherently rely on neural networks to handle high-dimensional data in the same way.

When the environment is high-dimensional, you can’t keep a separate value for every state-action pair in a table. Instead, you use a function approximator that can generalize across many states. Neural networks serve this role well, allowing the agent to estimate value functions from complex inputs. Deep Reinforcement Learning uses these networks to approximate the value function (or Q-function) directly from raw observations, rather than storing values in a tabular form. A common example is a network that takes the current state and outputs a Q-value for each possible action, guiding decisions through learning signals like TD errors. This approach scales to rich inputs such as images or high-dimensional sensor data, where tabular methods would be impractical. While other methods involve temporal-difference updates or Monte Carlo returns, they don’t inherently rely on neural networks to handle high-dimensional data in the same way.