HarvardX's Computer Science for Artificial Intelligence Professional Certificate

"Search" - A programme to determine how many degrees of separation there are between any two people in a dataset. In this project's example, I connected actors via movies they starred in, based on IMDB data. This programme returns the shortest path as a list of tuples describing the steps/ path. 

"Search" - In this problem-set I implemented python functions for optimal decision-making in a game of TicTacToe. The "AI" player is able to consider the current state of the game and evaluate all possible actions and counter-actions by an opponent by generating a utility value for all current and future moves. This is done following the minimax logic in which player A's goal is to minimise an outcome, while player B tries to maximise an outcome. Since implementing this, I have not been able to beat the programme (only tie or lose). 

"Knowledge" - This task required the implementation of different actor's knowledge bases as well as a logical model checking algorithm to determine "lies" and "truths".

"Knowledge" - This programme implements propositional logic in which an agent makes decisions by considering a knowledge base. The agent plays by making inferences and adding to/updating the knowledge base as the game progresses. On occasion, in minesweeper all "known" safe spots are exhausted and the agent must take a random chance - so this programme will not win every time, but it is able to consider all knowable information and make logical decisions. The only times it will fail is when it is forced to randomly choose and lands on a mine.

"Uncertainty" - In this project, I calculated a given page's PageRank, by:

(1) a random-surfer model to mimic the behaviour of a hypothetical surfer on the internet who clicks through pages. This model can be interpreted as a Markov Chain where a page represents a state, and each page has a transition model that chooses among its links at random.

​

(2) an iterative algorithm to define a page's PageRank using a recursive mathematical expression to calculate the probability that a random surfer will end up on a given page. PR(p) = ((1 - d) / N) + dΣi (PR(i) / NumLinks(i))

PR: page rank, d: damping factor, N: total number of pages in corpus, i: ranges over all pages that link to page p, NumLinks(i) is number of links present in page i.

"Uncertainty" - Here, I wrote functions that ultimately calculate probabilities that a given person has a genetic trait based on lineage, generic likelihoods, and whether or not a trait is observed in a person. Specifically, I wrote a function for joint probability calculation for every possible scenario given some facts about lineage and traits, and generated a probabilities dictionary to continuously update probabilities after each scenario and normalise the values to return a final probability distribution for each person on whether they have 0, 1, or 2 active genes and whether they do or don't show the trait.

"Language" -  In this project, I created a simple question answering algorithm by pre-processing a corpus of documents and computing Inverse Document Frequency scores, then processing a user's question using Term-Frequency Inverse-Document-Frequency and ranking top documents and top sentences. Although this algorithm is only simple compared to what is used in many applications today, I really enjoyed testing this programme after I completed it! 

"Learning" - Using Q-Learning, this programme learns through reward values for every possible (state, action) pair in a game of Nim. It learns both the value for a current state and the best future reward. After a set of 10,000 training games, the programme is unbeatable at the game.

"Optimisation" - Given a corpus of words and a crossword structure, this programme returns a complete crossword solution. It checks every word for each position and enforces node- and arc-consistency i.e. that it fits the the space and to its neighbours. It favours words that allow for a maximum of neighbouring words to arrive at an overall solution faster. The programme implements an AC3 algorithm, checks for completeness, checks for consistency, orders all words by how many other words it matches with in the range of possible words, and backtracks to try to fill empty spaces (amongst other functions).

​

██████████████

███████M████R█

█INTELLIGENCE█

█ N █████ N ████ S █

█ F ██L O G I C ███ O █

█ E █████ M ████ L █ 

█ R ███ S E A R C H █ V █

███████ X ████ E █ ██████████████

​

"Learning" - Given a range of features, this programme builds a nearest neighbour classifier to predict whether or not a new customer will complete an online purchase. The performance is measured in sensitivity and specificity.

"Neural Networks" -  In this project, I used TensorFlow for the first time to create a convolutional neural network to recognise road signs. 

"Language" -  In this project, I applied pre-processing steps to a given text (tokenising, stop word removal, lowercasing, punctuation removal), created a set of context-free grammar rules to parse a sentence and return an nltk tree that represents the syntax. The programme returns all possible "noun phrase" chunks.