Yоu аre trаining аn agent using a Deep Q-Netwоrk (DQN) in a discrete-actiоn gridworld. The agent can move up, down, left, or right. States are represented as one-hot vectors corresponding to positions on the grid. Although the agent achieves low training loss, its learned policy is unstable and performs poorly during evaluation episodes. Which of the following changes are likely to improve generalization and training stability? Select all that apply: