intrinsic motivation

**Intrinsic Motivation** in RL is the **use of internally generated reward signals to drive exploration** — augmenting external (task) rewards with intrinsic rewards based on novelty, curiosity, surprise, or competence, enabling the agent to explore effectively even without external reward. **Intrinsic Reward Types** - **Curiosity**: Reward for encountering states that are hard to predict — prediction error as reward. - **Count-Based**: Reward inversely proportional to visitation count — visit novel states. - **Information Gain**: Reward for actions that reduce uncertainty about the environment model. - **Empowerment**: Reward for states where the agent has maximum control over future outcomes. **Why It Matters** - **Sparse Rewards**: Many real-world tasks have extremely sparse external rewards — intrinsic motivation enables learning. - **Exploration**: Intrinsic rewards drive systematic exploration of the environment — avoids random wandering. - **Autonomy**: Enables agents to learn useful skills without any external reward — pre-training for downstream tasks. **Intrinsic Motivation** is **self-driven curiosity** — generating internal rewards to explore and learn even without external feedback.