SHAKYS.

Inspiration

Coming into the hackathon, we wanted to create a next-generation game that dynamically changes the behavior of our AI based on the player's past behaviors.

We created a difficult game by adjusting boss's behaviors to counter how a player decided to play. In this manner, we would be able to create a game that was unique to each player's playstyle and potentially each playthrough.

We created 2 entities, a boss and a player, each with unique movements and interactions. During a level, we logged player actions (move, attack, dash, etc.) and observable platform variables (entity velocities, distances, elevation, etc.) and used these as input nodes for a neural network, whose output nodes would control the boss's actions (move, attack, dash, etc.). To train the neural network, we wanted to implement a reward system where the boss would be rewarded for minimizing damage taken and maximizing damage output.

We found difficulty training the neural network because the data being used as inputs were dynamic to how they were rewarded (if near identical inputs were used, there would be scenarios where there would be completely opposite rewards). Hence, this made it almost impossible for the neural network to find accurate weights to come to an optimal decision on how to react and adjust to the player's behavior.

We had to change the scope and direction of our project multiple times and were able to persevere through the difficulties we encountered with the neural network (which was the unique selling point of the game). Despite not being able to successfully implement our neural network, we were able to create an adaptive AI instead which logged the different type of inputs by the player and behave responsively. We made a custom score for the game using FL Studios which we were all very proud of. Additionally, we are all very proud of ourselves for succesfully finishing and submitting in our first official hackathon.

We really expanded our knowledge and understanding of neural networks and how they operate. We also became aware of different strategies of other machine learning scientists might impelement such as evolution. Additionally, we learned that when managing projects and changing the scope of the end deliverable program, it is sometimes more efficient to start over from scratch rather than continuing to edit and debug what you already have.