Projects for udacity machine learning nanodegree
优达学城机器学习进阶项目

Titanic Survivor Prediction Built a simple decision tree classifier to predict passenger survival outcomes based on features such as gender and age from the Titanic disaster dataset, achieving an accuracy above 80%.

Boston Housing Price Prediction Utilized Python libraries like NumPy and Scikit-Learn to predict the selling prices of new homes using the Boston Housing dataset.

• Applied techniques such as K-Fold cross-validation and GridSearch to optimize machine learning model hyperparameters.
• Analyzed performance metrics and charts of various machine learning algorithms to evaluate the variance and accuracy of different models under various parameter settings.
• Selected the best-performing model, which achieved an R^2 score of 0.84 on the test set.

Finding Donors for CharityML Developed an efficient machine learning algorithm for CharityML, a Silicon Valley non-profit organization, to identify potential donors and reduce the cost of sending promotional emails.

• Preprocessed data using normalization, one-hot encoding, and other techniques.
• Trained and compared supervised learning models, including Gaussian Naive Bayes, SVM, and ensemble learning, ultimately selecting the top-performing model (ensemble learning with an F-score of 0.7645).
• Discussed the strengths, weaknesses, and characteristics of the three different models: Gaussian Naive Bayes, SVM, and ensemble learning.
• Evaluated feature importance using Sklearn's feature_importance_ attribute.

Customer Segmentation Analyzed data from a wholesale distributor in Lisbon, Portugal, to uncover hidden customer segments.

• Applied preprocessing techniques for feature correlation analysis, anomaly detection, and feature scaling.
• Employed Principal Component Analysis (PCA) for dimensionality reduction on the wholesale customer data.
• Utilized the Gaussian Mixture Model clustering algorithm (EM) to discover latent customer segments, revealing two primary groups: food service and supermarket customers, aligning with real-world observations.

Training a Smart Car to Drive Employed reinforcement learning techniques to train a self-driving car to efficiently navigate and reach its destination within a constrained time frame in a simplified environment.

• Determined the state space of the smart car by evaluating safety and reliability metrics while considering the total number of states (computational complexity) for optimization.
• Implemented the Q-Learning algorithm to train the AI agent to make optimal decisions based on its environment.
• Fine-tuned the exploration factor decay function, learning rate, and other parameters, enabling the smart car to achieve A+ ratings in both safety and reliability.
• Discussed the ineffectiveness of the future reward gamma in this context, considering that the traffic problem at hand is not a Markov Decision Process (MDP).

Version 1.1 | Update Date: 2018/1/31 | Update Content: Added concise descriptions for each project and included links.