|
|
Optimizing Super Mario game tree search
Šosvald, David ; Gemrot, Jakub (advisor) ; Dvořák, Tomáš (referee)
Super Mario Bros. is still actively used as a model game for research in level genera- tion. Every year, the most recent techniques are applied and tested. This lately includes various deep learning and reinforcement learning methods. Many of the level generators use an artificial agent to test levels' playability or to gather playthrough metrics. There- fore, the performance of the level generators is undeniably tied to the performance of the artificial agent used, both in level validation and the computing time needed. In our previous work, we created a new state-of-the-art agent for Super Mario Bros. as a proof of concept when we implemented a more efficient forward model (world simu- lation) for the Mario AI framework. In this work, we continue in that work and focus on optimising how the agents explore the game tree by devising domain-specific heuristics and running extensive parameter searches to tune the agents as much as possible. Thanks to these improvements, a new state-of-the-art agent was created. This new agent should be capable of beating every standard Super Mario Bros. level and it requires less time to solve levels than previous agents. We also present a proof of concept agent that is capable of solving maze-like levels, which is something none of the previous agents was capable of.
|
|
|
Efficient forward model for Super Mario AI framework
Šosvald, David ; Gemrot, Jakub (advisor) ; Dingle, Adam (referee)
The artificial intelligence framework for Super Mario Bros. has been used in a lot of research in the past decade. We have noticed that the forward model (simulation of the game world) present in the framework is quite inefficient and thus negatively influences all work that is based on it, especially intelligent agents that utilize it. This means that every work using such agents is influenced too. That might also include level generation, where agents are often used to test the playability and properties of levels. We have implemented a more efficient forward model and as a proof of concept, we used the improved forward model to create new intelligent agents. In a benchmark we ran, the negative influence of the original forward model was confirmed, as all agents using the new forward model greatly outperformed agents based on the original one.
|
guest ::
