What is intrinsic motivation?

What is intrinsic motivation?

Intrinsic motivation is a concept from psychology, developed in self-determination theory by Edward Deci and Richard Ryan, describing behavior pursued for its own inherent satisfaction rather than for an external reward. In reinforcement learning, the term has been adapted to describe agents that generate their own internal reward signals, such as curiosity or novelty, to guide learning and exploration when external feedback is sparse or absent. Unlike extrinsic motivation, which relies on external rewards or punishments, intrinsic motivation is driven by factors internal to the agent and can sustain learning over time.

What are some common causes of intrinsic motivation in psychology?

Self-determination theory identifies several drivers of intrinsic motivation in people:

• Personal interest: Having a genuine curiosity or passion for the topic or activity at hand.
• Autonomy: Having a sense of control, choice, and ownership over one's actions and decisions.
• Competence: Having a belief in one's ability to master the skills and knowledge required for the task or challenge.
• Relatedness: Having a sense of connection, belonging, and support from others who share similar goals or values.

How is intrinsic motivation used in artificial intelligence?

In AI, intrinsic motivation most commonly appears in reinforcement learning, where an agent is given an internally generated reward—such as curiosity about unfamiliar states, novelty, or prediction error—to guide exploration when external rewards are sparse or delayed. This helps agents explore more effectively and acquire skills that are useful even before an explicit task reward is available.

What are some benefits of using intrinsic motivation in AI?

Intrinsic reward signals help reinforcement learning agents explore more effectively in environments where external rewards are sparse, delayed, or absent. This can speed up learning and lead agents to acquire skills or knowledge that later prove useful for a range of downstream tasks, rather than optimizing narrowly for a single external reward.

What are some challenges associated with using intrinsic motivation in AI?

Designing effective intrinsic reward signals is difficult: measures like novelty or prediction error can be gamed by an agent seeking out unpredictable or noisy parts of an environment rather than genuinely useful experiences. Intrinsic rewards can also compete with or distort the external task reward if not carefully balanced, making the resulting behavior harder to predict, evaluate, and compare across systems.