Reinforcement-learning lets algorithms reprogram themselves using internal performance and reward subroutines.

Evidence

Reinforcement-learning was the child of software engineers who realised that algorithms had the potential to evaluate their own performances – and make improvements – far faster than any human could. To achieve this, they wrote into them two types of subprograms (or subroutines): one that measures the algorithm’s performance while it is in action and at tremendous speeds, and another (called a reward function) that helps the algorithm alter itself so as to improve its performance in accordance with the engineers’ objectives. — Yanis Varoufakis, Technofeudalism, 311

Categorization

• DEFINE: Defines the technical mechanism of reinforcement-learning in accessible terms: performance-measurement plus reward function subroutines.
• XREF: connects to: Reinforcement-learning lets algorithms reshape themselves based on how they influence humans. — Yanis Varoufakis, Technofeudalism: What Killed Capitalism, 312; Reinforcement-learning lets algorithms improve themselves faster than humans can.; Machine learning divides into unsupervised, supervised, and reinforcement learning — Brian Christian, The Alignment Problem, 58 (+4 more)

Broader Connections

• Machine learning divides into unsupervised, supervised, and reinforcement learning #exemplifies (0.97) — Brian Christian, The Alignment Problem, 58
• Reinforcement-learning lets algorithms improve themselves faster than humans can. #expands (0.95)
• Economies of action get institutionalized into automated algorithms that modify behavior #expands (0.90) — Shoshana Zuboff, The Age of Surveillance Capitalism, 939
• AI systems can fail through instrumentalized self-preservation goals #complicates (0.80) — , The Braided Reading EPUB, 1604
• Complex behavior emerges from rewarding successive small approximations #corroborates (0.72) — Brian Christian, The Alignment Problem, 660
• Systems can learn from shifting their own expectations before final outcomes arrive #expands (0.72) — Brian Christian, The Alignment Problem, 600