A Comparative Study of Various Artificial Intelligence Based Agents for the Game of Angry Birds with and without Splitting

Abstract

In a game of angry birds, birds are fired from a slingshot and are targeted towards stationary pigs located at different fixed distances from the slingshot. The angry birds have to be fired in such a way that it lands as close as possible to the pigs' location. The goal is to develop an artificial intelligence-based model that would play the angry birds game based on the past human experience. In this game, the user will give the initial velocity and the angle of projection. Based on these parameters, the shot will be played, and the outcome is stored as a tuple consisting of the initial velocity, the angle of projection, and the location of pigs that have not been destroyed in a database. The machine learning-based agent reads the data from the database, trains itself based on the outcome of previous shots stored in the database, and plays the best possible shot according to the data retrieved from the database. Two machine learning models have been proposed, which are the K Nearest Neighbours model and the Naive Bayes model. The third model is the stochastic gradient descent model, which plays a shot based on the minimization of the distance between the angry bird and the pig using an objective function in terms of the initial velocity and splitting angle. The performance of both these agents has been compared with the human agent's performance in terms of the average number of wins per 100 games.