Which method chooses model parameters so predictions are as close as possible to actual data on average?

Which method chooses model parameters so predictions are as close as possible to actual data on average?

• A. Policy Gradient approach
• B. Residual Sum of Squares (RSS)
• C. Ordinary Least Squares (OLS)
• D. Maximum Likelihood Estimation (MLE)

Answer: (C)

The idea is to pick model parameters that directly minimize how far predictions are from actual observations on average. In linear regression, predictions are ŷ = Xβ. Ordinary Least Squares finds the β that minimizes the sum of squared residuals, ∑(y_i − ŷ_i)^2, which is the mean squared error up to a constant factor. By continually reducing these squared differences across all data points, OLS yields parameter estimates that make the predictions as close as possible to the real data on average. Under standard assumptions, these estimates are unbiased and efficient among linear estimators. Policy gradient is used to optimize decision policies in reinforcement learning, not to fit regression models. Residual Sum of Squares is the objective that OLS minimizes, not a separate estimation method. Maximum Likelihood Estimation is a broader framework and, for normally distributed errors in a linear model, ends up giving the same estimates as OLS, but the direct approach described here is the least-squares one.

The idea is to pick model parameters that directly minimize how far predictions are from actual observations on average. In linear regression, predictions are ŷ = Xβ. Ordinary Least Squares finds the β that minimizes the sum of squared residuals, ∑(y_i − ŷ_i)^2, which is the mean squared error up to a constant factor. By continually reducing these squared differences across all data points, OLS yields parameter estimates that make the predictions as close as possible to the real data on average. Under standard assumptions, these estimates are unbiased and efficient among linear estimators.

Policy gradient is used to optimize decision policies in reinforcement learning, not to fit regression models. Residual Sum of Squares is the objective that OLS minimizes, not a separate estimation method. Maximum Likelihood Estimation is a broader framework and, for normally distributed errors in a linear model, ends up giving the same estimates as OLS, but the direct approach described here is the least-squares one.