A4.3 Machine learning approaches (HL only)

The question asks about the policy component in a reinforcement learning system. We need to recall the precise definition of policy from the RL framework.

In reinforcement learning, a policy is defined as the strategy that maps states to actions. It tells the agent what action to take given the current state of the environment.

For the warehouse robot, the policy would encode decisions such as: given that the robot is at grid position (3,5) with a clear path ahead, move forward. This is exactly 'the learned mapping from the robot's current situation to the action it should take'.

The correct answer is 'The learned mapping from the robot's current situation to the action it should take', which precisely matches the definition of policy as a state-to-action mapping.

We need to identify which option correctly defines the policy in reinforcement learning, applied to a warehouse robot scenario.

'The numerical score assigned to the robot after each successful delivery' describes the reward, not the policy. Rewards are feedback signals from the environment, not decision-making strategies.

'The physical layout of the warehouse including shelves and corridors' describes the environment (or aspects of the state space). The environment is the context in which the agent operates, not the agent's decision strategy.

'The total accumulated reward the robot has earned since training began' describes the cumulative reward, which is the objective the agent tries to maximise — not the strategy it uses to make decisions.

By elimination, 'The learned mapping from the robot's current situation to the action it should take' is correct. This matches the definition of policy as a mapping from states to actions.

The model is $\text{JourneyTime}=4.2+0.35\times \text{Signals}$, where ${\beta }_0=4.2$ (intercept) and ${\beta }_1=0.35$ (slope). We need to predict the value when Signals $=20$.

$\text{JourneyTime}=4.2+0.35\times 20$

$\text{JourneyTime}=4.2+7.0=11.2\text{ minutes}$

The predicted journey time with 20 active signals is 11.2 minutes.

We must substitute Signals $=20$ into the regression equation and compute the result. We can verify each option by checking if it matches the correct calculation.

7.0 minutes would result from computing $0.35\times 20=7.0$ but forgetting to add the intercept ${\beta }_0=4.2$. This is a common error — the intercept must always be included.

9.7 minutes might arise from an arithmetic error, such as computing $4.2+0.35\times 16=9.8$ or a similar miscalculation. It does not result from correctly applying the given equation with Signals $=20$.

8.4 minutes could result from incorrectly using $0.35\times 12+4.2=8.4$ or confusing the coefficient. It does not match $4.2+0.35\times 20$.

The correct calculation gives $4.2+7.0=11.2$ minutes, confirming this is the correct answer.