ML unit-4 notes

https://drive.google.com/file/d/199h19B7eV-ANGz94jmEzJvL97xDBgCuW/view?usp=sharing

This document provides an overview of several machine learning topics, including Genetic Algorithms (GAs), Genetic Programming (GP), Learning Sets of Rules, Sequential Covering Algorithms, FOIL Algorithm, Inverse Resolution, and Reinforcement Learning (RL). Genetic Algorithms, inspired by biological systems, search a space of candidate hypotheses, often represented by bit strings, using operations like mutation and crossover to identify the best hypothesis based on a fitness value. Genetic Programming extends GAs where the individuals in the population are computer programs represented by trees. Learning sets of rules, especially "first-order Horn clauses" (FOHCs), is discussed, with Sequential Covering Algorithms like CN2 performing a greedy, general-to-specific search to learn rules one at a time. The FOIL algorithm is a specific approach for learning first-order rules, also employing a general-to-specific hill climbing search within its inner loop. Finally, Reinforcement Learning is introduced, where an agent learns to choose optimal actions in an environment to maximize cumulative reward, often using a policy and the Q-learning algorithm.

Key Topics and Definitions:

• Genetic Algorithms (GA): Algorithms that generate successor hypotheses by repeatedly mutating and recombining parts of the best currently known hypotheses, often represented by bit strings, to search a space and identify the best hypothesis based on fitness.
• Genetic Programming (GP): A form of computation where individuals in the population are computer programs, typically represented by trees, rather than bit strings.
• Sequential Covering Algorithms: Algorithms that learn sets of rules by learning one rule, removing the data it covers, and then iterating this process.
• FOIL Algorithm: An algorithm for learning sets of first-order rules (similar to Horn clauses), which uses a general-to-specific hill climbing search to specialize rules until they avoid all negative examples.
• Q-Learning: A Reinforcement Learning algorithm used by an agent to learn a policy for selecting its next action based on the current observed state, which involves updating a table entry for an evaluation function $\hat{Q}(s,a)$.