11[
2+ " About this project"
3+ " "
24 " ----------------------------------------------------------------------------------------------------"
35 " This project involves using model-based reinforcement learning to solve the 15-puzzle problem, "
46 " a sliding puzzle where the goal is to arrange numbered tiles in the correct order. The approach "
57 " relies on a Q-Table, a tabular data structure that stores values representing the expected utility of "
68 " taking specific actions in each possible state. The Q-Table helps the learning agent make decisions "
79 " by providing a way to evaluate which actions lead to more favorable outcomes over time."
8- " "
9- " In this implementation, key parameters like learning rate and discount factor control how the "
10- " agent updates its understanding of the environment. The learning rate (learningRate in "
11- " config.json) determines how much new information overrides old knowledge, while the "
12- " discount factor (discount) influences how future rewards are valued relative to immediate ones."
13- " The configuration allows for the generation of a specified number of training examples (or "
14- " 'lessons') to teach the model through repeated trials. For each lesson, various parameters such as "
15- " goals, start positions, and locked elements are defined, guiding the agent’s training process."
16- " The settings in config.json define these learning conditions, providing a structured approach to "
17- " gradually mastering the puzzle."
1810 " ----------------------------------------------------------------------------------------------------"
1911 " "
20- " about puzzle - https://en.wikipedia.org/wiki/15_puzzle"
21- " project source code - https://github.com/zc42/ml-puzzle-15-v1"
22- " java port - https://github.com/zc42/game_15_java/blob/main/README.md"
23- " python port for kaggle - https://www.kaggle.com/code/zilvinasc/reinforced-learning-qtable-puzzle-15"
24- " development history - https://github.com/zc42/stuff"
12+ " About puzzle - https://en.wikipedia.org/wiki/15_puzzle"
13+ " Project source code - https://github.com/zc42/ml-puzzle-15-v1"
14+ " Java port - https://github.com/zc42/game_15_java/blob/main/README.md"
15+ " Python port for kaggle - https://www.kaggle.com/code/zilvinasc/reinforced-learning-qtable-puzzle-15"
16+ " Development history - https://github.com/zc42/stuff"
2517 " ----------------------------------------------------------------------------------------------------"
2618 " "
2719 " "
28- " CONFIGURATION HELP"
29- " "
30- " "
31- " Configuration is editable."
20+ " Project configuration properties (configuration is editable)"
3221 " ----------------------------------------------------------------------------------------------------"
3322 " usePretrainedDataWhileTesting - boolean, true or false."
3423 " When set to false, the q-table data learned in this session will be used."
4938 " startPositions - number array, starting positions for the empty tile."
5039 " lockedElements - number array, tiles that are fixed and cannot be moved, usually representing the goals from previous lessons."
5140 " All number arrays, valid numbers are from 1 to 15, movable tile nubers in a puzzle."
52- " "
41+ " ------------------- "
5342 " lessonsToGenerate - number, the number of episodes that will be generated for this particular lesson in each training cycle."
5443 " ----------------------------------------------------------------------------------------------------"
55- " ---------------------------------------------------------------------------------------------------- "
44+ " "
5645 " "
5746 " Some theory"
5847 " ----------------------------------------------------------------------------------------------------"
7766 " The Q-Table helps an agent learn which actions to take in each state to maximize its cumulative reward over time."
7867 " By iteratively updating the Q-values based on experiences, the agent learns an optimal policy for decision-making."
7968 " ----------------------------------------------------------------------------------------------------"
80- " "
8169 " Model-based reinforcement learning"
8270 " "
8371 " Model-based reinforcement learning is an approach where the agent learns a model of the "
8472 " environment, including how actions affect states and the rewards they yield. The agent uses "
8573 " this model to simulate future outcomes and plan its actions, making decisions based on predictions "
86- " rather than just past experiences. This can improve learning efficiency since the agent can learn "
87- " from simulated interactions, but it depends on the accuracy of the learned model."
74+ " rather than just past experiences. This can improve learning efficiency since the agent can learn "
75+ " from simulated interactions, but it depends on the accuracy of the learned model."
8876]
0 commit comments