Learn Before
Goal of Reinforcement Learning
The primary objective in reinforcement learning is to develop a policy that enables an agent to maximize the total cumulative reward, also known as the return, that it accumulates over an extended period of interaction with its environment.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Agent Performance Calculation
An agent interacts with an environment over a sequence of four time steps. The rewards it receives at each step are as follows: r₁ = +3, r₂ = -1, r₃ = +5, r₄ = -2. What is the total cumulative reward for this entire sequence?
Consider an agent that completes a five-step sequence of actions, receiving the following rewards at each step: [-5, +1, +1, +1, 0]. This sequence is preferable to another sequence that consists of a single step with a reward of -1.
Learn After
Objective Function as Expected Cumulative Reward (Performance Function)
An agent is being trained to find the best route through a system. It is presented with two options:
- Route 1: Provides a consistent, small positive reward at every step, resulting in a total reward of +15 for the entire route.
- Route 2: Starts with a step that gives a negative reward (a penalty) of -5, but subsequent steps lead to very high rewards, resulting in a total reward of +50 for the entire route.
An agent that has been successfully trained according to the primary objec
Analysis of a Suboptimal Agent Policy
An agent is learning to play a game where the objective is to get the highest possible final score. At a critical decision point, the agent chooses an action that yields an immediate reward of 0, passing up an alternative action that would have given an immediate reward of +10. This decision is necessarily an indication that the agent's policy is flawed and not aligned with the primary goal of its learning framework.