Concepts / Exploration vs Exploitation

Exploration vs Exploitation

Greedy actions are those with the greatest current estimated value.

  • Machine Learning

One Choice, Two Purposes

When an agent must choose an action, it has two different ways to use its current situation. It can rely on what it currently believes about the actions and seek the best expected reward for the next step. Or it can choose an action to learn more about it. These purposes create the exploration-versus-exploitation problem.

A single action selection cannot both explore and exploit. Selecting the currently highest-valued action pursues immediate reward; selecting a different action creates an opportunity to improve knowledge about that action.

uses current knowledgegathers informationBest-known actioncurrent highest estimateImmediate rewardNongreedy actioninformation opportunityImproved estimate
What is the difference between using the best-known action and trying an action whose value is uncertain?

Finding the Greedy Action

The agent has an estimated value for each available action. The greedy action is the action with the greatest current estimated value. “Greedy” describes the choice at the present decision, not a guarantee that the action will always produce the greatest reward.

Comparing three estimated action values

Suppose an agent currently estimates the values of three actions as A = 4, B = 7, and C = 5. Which action is greedy?

List the estimates: The current estimates are 4 for A, 7 for B, and 5 for C.

Locate the greatest estimate: The greatest current estimated value is 7.

Identify the action: Action B has the greatest estimated value, so B is the greedy action in this situation.

Action B is greedy because its current estimated value is greatest.

maximum valueAction Aestimate 4Action Bgreatest estimateAction Bestimate 7Action Cestimate 5
Given several actions and their estimated values, which action is greedy, and how does the maximum value determine the choice?

Exploitation and Exploration

Exploitation means choosing the greedy action: the agent uses its current knowledge to pursue the best expected one-step reward. Exploration means choosing a nongreedy action: the agent gathers information that may improve its estimates later.

ChoiceAction selectedPrimary purpose
ExploitationGreedy actionPursue the best expected one-step reward
ExplorationNongreedy actionGather information that may improve later estimates

The distinction is based on both the selected action and the purpose of the choice.

inspect current knowledgeseek immediate rewardlearn moreCurrent estimatesactions have estimatedvaluesChoice purposeimmediate reward orlearningGreedy actionexploitationNongreedy actionexploration
How does an agent decide whether to choose the currently highest-valued action or a different action for exploration?

Short-Term and Long-Term Outcomes

Exploitation can be the best choice for the next step because it uses the action with the greatest current estimated value. Exploration can be worthwhile when many future selections remain: trying a nongreedy action may reveal that it is better than currently estimated, allowing improved choices later.

A short-term sacrifice for later knowledge

An agent estimates that Action A is better than Action B. It can choose A now for the best expected one-step reward, or choose B to learn more about B. How do the two choices differ?

Exploit now: Choosing A follows the current estimate and aims at the best expected immediate reward.

Explore now: Choosing B is nongreedy under the current estimates, but it creates an opportunity to improve the estimate for B.

Consider future selections: If many future selections remain, information from B may support better choices later and may produce greater total reward.

Exploitation favors the immediate decision; exploration may improve the quality of later decisions.

use current estimatenext stepseek informationfuture selectionsAction selectionGreedy actionbest expected one-steprewardImmediate rewardNongreedy actionlearn moreLater choicespossibly better totalreward
How can exploiting the best-known action produce a higher immediate reward while exploring another action lead to better rewards later?

Common Reasoning Errors

  • Treating the greedy action as the action that is guaranteed to be best.

    A greedy action is defined by having the greatest current estimated value, not by a guarantee about its actual reward.

    Fix: Describe it as the action currently estimated to have the greatest value.

  • Calling every nongreedy choice irrational.

    A nongreedy choice can be exploration, which may reveal a better action and improve later estimates.

    Fix: Ask whether the choice is intended to gather information for future decisions.

  • Assuming exploration is always better than exploitation.

    Exploration and exploitation serve different purposes and horizons.

    Fix: Separate the immediate-reward goal from the possible long-term value of improved knowledge.

  • Claiming that one action selection both explores and exploits.

    Selecting the greedy action uses the current estimate for immediate reward, whereas selecting a nongreedy action creates an opportunity to learn about that action.

    Fix: Classify the selected action as exploitation if it is greedy and exploration if it is nongreedy.

Practice the Decision

EASY

An agent has current estimated values of 3 for Action A, 6 for Action B, and 4 for Action C. Identify the greedy action. Then describe what it means to exploit in this situation and what it means to explore instead.

Hints
  • Start by finding the greatest current estimated value.
  • Exploitation selects the action with that greatest estimate.
  • Exploration selects a nongreedy action to gather information that may improve later estimates.

A complete explanation should name the greedy action, connect it to exploitation, and explain that choosing another action would be exploration because it prioritizes information rather than the best expected one-step reward.

Key Takeaways

  • A greedy action has the greatest current estimated value.
  • Exploitation chooses the greedy action to pursue the best expected one-step reward.
  • Exploration chooses a nongreedy action to gather information that may improve later estimates.
  • Exploitation can favor immediate reward, while exploration can improve total reward when many future selections remain.
  • One action selection cannot simultaneously serve both exploration and exploitation.