Epsilon-Greedy Algorithm

The mechanism separates repeated decisions into two modes. Exploitation collects value from what currently looks best; exploration samples alternatives that might overturn that ranking. Epsilon fixes the probability of exploring, so success cannot quietly eliminate experimentation. Unlike an epsilon-decreasing strategy, the discovery budget does not automatically shrink with time.

That standing budget matters because “best-known” is not the same as best. Your observations may be incomplete, and a current winner can remain unchallenged simply because competing options receive no further trials. Random exploration keeps producing information, while the larger exploit share prevents learning from consuming the whole decision process.

Epsilon-greedy fixes the exploration rate instead of adapting it to consequences or accumulated knowledge. Twenty percent may be too timid when little is known, wasteful after the field is well understood, or dangerous when a random choice has serious downside. The mechanism supplies discipline, not the correct epsilon; that still depends on the cost of exploration.