Adriana Haddad

Beginnings of an experiment:

I had the pleasure of working with a talented team at the University of Michigan to create a game centered around the different museum collections the university had to offer. Collect Connect was meant to inspire creative thinking between the players and spark conversations about players’ schemas and associations between varying museum pieces. We wanted to take these very human thought patterns and use them to train an AI to be able to follow these seemingly random jumps in thinking and create its own associations between objects. We did not want to train an AI to identify objects or simply their physical/ functional capabilities, but rather to attempt to move to a more abstract interpretation between objects.

For a majority of my time on the Collect Connect team I was the solo gameplay programmer for the 2 player version of the project. During the development cycle I worked on creating rapid prototypes for the latest version of the game, a majority of the time using placeholder assets to test varying look and feel of the game. These prototypes helped lay the groundwork for the team to decide the final direction on game play and style. While working on this project, I was very careful with the underlying design, especially in the early stages where gameplay wasn’t locked in and pivoted often. I wanted to make sure the code base was scaleable and easy to maintain as features needed to be added; I felt especially accomplished on this aspect when in the final hours before we launched the game, the designers requested some changes and additions to the game flow and I was easily able to implement those request without worrying if they were going to negatively impact another aspect of the game. That was probably the biggest miracle of last minute changes which actually did what they were supposed to do and didn't break anything along the way.

I’m still shooketh.

Alpha/ Beta Pruning:

Taxman is a game whose goal is to show the effectiveness of min maxing and how alpha beta pruning can expedite those results. It is a single player nim game created in Unity where players face off against the ever calculating taxman. The goal of the game is to accrue more money than the opponent, but with every choice a player makes there is the potential for their opponent to also gain points.

The taxman is a tough opponent as it searches hundreds of thousands of potential possibilities in advance depending on the players set preferences. Players are able to adjust the depth at which the taxman searches, whether it uses alpha/ beta pruning, and adjust the turn length of the taxman’s turn.

This project was a simple one to create game play wise, but it was the first time I needed to use multithreading in Unity. Min maxing is quite a resource intensive search, and in order for the taxman's turn to not clog up the main thread it needed to be moved to its own. This move also allowed the game to have the real time time limit for taxman's turn, otherwise players would have to estimate how patient they’re willing to be before his turn starts. This gave me the opportunity to learn about the capabilities of Unity’s multithreading system and a chance to test it out.

Part of a series of tools created for Dr. Maxim for use in his artificial intelligence class.

Heuristic Search Tool:

Pathfinder is an app made to display and explain the effectiveness of different heuristic searches. It displays Dijkstra,Breadth First, Depth First, Branch and Bound, A*, Hill Climbing, and Best First. It shows all explored paths and the accepted path, giving total distance traveled so users can more easily compare efficiency. Users can customize start and goal locations as well as adding any barriers or difficult terrain(squares that are traversable but cost more than a normal square) to test interesting situations. They can also play around with the underestimators and the weight it has in calculating distance.

Part of a series of tools created for Dr. Maxim for use in his artificial intelligence class.

Neural Network with Reinforcement Learning:

This project was created using Unity’s ML agents and contains multiple pretrained models to compare. It allows users to immediately see the effects of varying the weights the neural network had been trained with on the behaviour. Since nn can take a very long long time to train, we wanted the users to be able to experiment without the wait.

Users can see how raising/ lowering the rewards given to the ai can dramatically change the ai’s ability to complete the desired task. This also helps demonstrate a classic issue with training ai, they tend to cheat. I encountered this problem when initially setting up the rewards for aiming at a target. While developing I would set some parameters, sit back, and feel accomplished. And then I started watching the behaviour develop and realize the fool I’ve become. I had given the bot accumulating rewards for staying aimed at a target. That reward minus the time penalty was greater than the total reward for quickly shooting the targets.

This reinforcement learning project was a great way for me to jump in and start hands on experimenting with neural networks. There were a lot of learning curves, one such being credit assignment. For example, my first goal was to train the bot to accurately shoot a target.When the bot hit a target, the reward would be given the moment the bullet hits the target. If the bullet missed and hit a wall, they would get a negative reward. This set up caused a lot of random shots at best, and when the negative reward was too high, no shots at all. The ai had no way of crediting a positive shot to hitting a target and a negative one to missing. As a human who set this up I thought it was very obvious, until I read this article on training an ai to play google’s dino run. I had run into a credit assignment issue, since I was still learning I didn’t have the background knowledge to recognize at a glance. I was able to overcome this by breaking up shooting into aiming at a target and pulling the trigger. It cuts the corner a little, but the ai’s aren't the only ones who can cheat the system.

Part of a series of tools created for Dr. Maxim for use in his artificial intelligence class.

Steampunk Airship Senior Design Project

“Is a multiplayer tabletop turn-based game made in Unity for Android tablets to entertain users. Before the game is started, players will be allowed to tune settings relating to game difficulty, length, and various gameplay elements. Some of these settings include the number of players, the number of target objects needed for a player to win, enemy aggression levels, day and night cycle lengths, weather severity, and turn timers.

Core gameplay features include resource management, evasion of enemies, navigation around terrain obstacles, and racing against opponents. Players of the game will be encouraged to strategize in order to finish the game faster than their opponents. If a player feels they require more information during their turn, they will be able to toggle a hint button to help them. Also, if players wish to stop playing the game at any time, they have the option to save their game to be able to resume play in the future.”