Unleashing the Power of AI in Mario Speedruns

Table of Contents

1. Introduction
2. Creating the Mario AI
   • Remaking Super Mario Bros from scratch
   • Incorporating AI magic
   • Implementing proximal policy optimization (PPO)
   • Rewarding Mario's velocity
3. The Training Process
   • Teaching Mario through rewards and punishments
   • Iterative training and improvements
   • Overcoming limitations and discovering shortcuts
   • Analyzing AI's performance against speedrun records
4. Challenges and Limitations
   • Financial constraints and GPU usage
   • Selecting levels without shortcuts
   • Debugging code and fixing errors
5. Achieving Speedrun Records
   • Training the AI on level 5-2
   • Continuous improvements and iterations
   • Breaking speedrun records
6. Conclusion

Creating the World's Fastest Super Mario Bros Speedrun

Introduction

In the world of gaming, speedrunning has become a popular phenomenon. Gamers aim to complete a game as quickly as possible, often breaking records and challenging their own skills. Super Mario Bros, an iconic game loved by millions, has also attracted the attention of speedrunners. But what if we could create an artificial intelligence (AI) that could beat speedrun records in Super Mario Bros? In this article, we will delve into the journey of creating the world's fastest Super Mario Bros speedrun using AI technology.

Creating the Mario AI

Remaking Super Mario Bros from Scratch

The first step in our quest was to recreate the Super Mario Bros game environment. Initially, we began by watching tutorials and gathering assets to build the game from scratch. However, to our surprise, we discovered that an open-source AI environment called "goons open AI" had already provided a pre-made Mario environment. This saved us time by eliminating the need to recreate the entire game, allowing us to focus directly on incorporating AI elements.

Incorporating AI Magic

Once we had the Mario environment, we began integrating the AI magic. Our algorithm of choice was proximal policy optimization (PPO). While PPO sounds complex, it is merely a technique used to train AI models. Think of PPO as a parent and the Mario AI as a child. The parent rewards or punishes the child based on its actions, shaping its decision-making algorithm. Our goal was to train Mario to become a skilled Goomba eliminator.

Implementing Proximal Policy Optimization (PPO)

In our implementation of PPO, Mario received rewards for moving towards the right side of the level, as the flag indicating the end of each level is always on the right. Additionally, we punished Mario for dying and staying alive too long. This incentivized him to finish the level as quickly as possible, thus optimizing his speed. These tweaks were made through several iterations of training, constantly refining Mario's decision-making process and improving his performance.

The Training Process

Teaching Mario through Rewards and Punishments

The training process involved teaching Mario through a series of rewards and punishments. Whenever Mario made progress or achieved something positive, he received rewards, triggering a rush of dopamine in his virtual brain. Conversely, whenever Mario made mistakes or failed, he received punishments. This mechanism helped Mario mold his decision-making algorithm, aiming for faster completion of levels.

Iterative Training and Improvements

Training an AI model takes time and patience. We started with just five iterations of training, which allowed Mario to grasp the basics of the game. However, we knew that to achieve speedrun records, more training was required. Doubling the number of iterations, we observed significant improvements in Mario's performance. He learned to surpass the obstacles encountered in the early stages of the game and developed better strategies to move forward.

Overcoming Limitations and Discovering Shortcuts

During the training process, we faced several challenges and limitations. One particularly significant constraint was financial, as each training session required substantial GPU usage, causing budget constraints. Additionally, we encountered obstacles in selecting levels without shortcuts, to ensure Mario's performance was not influenced by skipping sections of the game. However, despite these limitations, we made progress, and our AI continued to learn and improve.

Analyzing AI's Performance against Speedrun Records

To assess our AI's progress, we compared its performance against existing speedrun records. In the initial stages, the AI fell short, but with persistent training, it closed the gap. Analyzing the AI's performance revealed areas where it outperformed human speedrunners, making us hopeful for even greater achievements.

Challenges and Limitations

Financial Constraints and GPU Usage

Training an AI model demands computational resources, and maintaining a high-performance GPU comes at a cost. Our budget limitations posed a challenge, forcing us to carefully allocate resources and select specific training sessions. However, despite these constraints, we made the best use of available resources to continue pushing the boundaries of Mario's speedrun abilities.

Selecting Levels without Shortcuts

To ensure fairness, we aimed to train Mario on levels without shortcuts. While these levels might not offer the fastest routes, it allowed us to evaluate the AI's true ability. By eliminating shortcuts, we aimed to achieve authentic speedrun records that truly showcased Mario's skills.

Debugging Code and Fixing Errors

During the training process, we encountered various bugs and errors in the code. Debugging and fixing these issues was a crucial part of refining our AI model. Resolving these challenges ensured that Mario's performance was not hindered by technical glitches and contributed to his continuous improvement.

Achieving Speedrun Records

Training the AI on Level 5-2

In our pursuit of breaking speedrun records, we trained our AI model extensively on level 5-2. By iterating and refining the training process, we observed remarkable progress. The AI learned to navigate the level swiftly, surpassing obstacles and showcasing its potential to achieve remarkable speed.

Continuous Improvements and Iterations

With each training session, our AI model continued to enhance its performance. Adjusting hyperparameters, investing additional hours of training, and fine-tuning the algorithms propelled Mario closer to speedrun records. The AI's ability to adapt and improve amazed us, highlighting the potential of machine learning in the gaming world.

Breaking Speedrun Records

After rigorous training and improvements, we witnessed a significant breakthrough. Our AI achieved a remarkable time of 3 minutes and 47 seconds on level 5-2, breaking the existing speedrun record. The joy and sense of accomplishment were unparalleled, as our creation surpassed human limitations and set a new standard in Super Mario Bros speedrunning.

Conclusion

In this journey of creating the world's fastest Super Mario Bros speedrun using AI, we encountered challenges, limitations, and moments of triumph. Through meticulous training, refining algorithms, and overcoming financial constraints, we were able to witness Mario's transformation into a speedrunning champion. This achievement highlights the tremendous potential of AI in the gaming world, pushing the boundaries of what is possible. The adventure continues as we strive to break more records and explore the endless possibilities that AI technology offers in the realm of gaming.

Highlights

• Building an AI-powered Super Mario Bros speedrun
• Incorporating proximal policy optimization (PPO) in the training process
• Rewarding Mario's velocity and optimizing his decision-making algorithm
• Overcoming challenges and budget constraints in training the AI
• Training the AI to surpass human speedrun records

FAQ

Q: How does the AI learn from rewards and punishments? A: The AI learns through a reward system, where positive actions are rewarded with dopamine rushes, and negative actions are punished. This shaping process helps the AI refine its decision-making algorithm.

Q: How did you ensure fairness in achieving speedrun records? A: We carefully selected levels without shortcuts to evaluate Mario's true abilities. This approach resulted in authentic speedrun records, showcasing Mario's skills without exploiting the game mechanics.

Q: What challenges did you face during the training process? A: One major challenge was financial constraints and the cost of maintaining high-performance GPUs for training. Additionally, debugging code and fixing errors were essential tasks to ensure smooth progress.

Q: How did you measure the AI's performance against world records? A: We compared the AI's completion time to existing speedrun records for specific levels. This provided a clear indication of Mario's progress and the extent to which he approached or surpassed world records.

Q: Are there any limitations to the AI's speedrun abilities? A: While the AI has made remarkable progress, there are certain limitations. For example, the AI may not be able to perform certain skips or tricks that require additional input, as it significantly affects training time.

Q: What are the future prospects for AI in gaming? A: The achievement of creating an AI-powered speedrun in Super Mario Bros opens up exciting possibilities for the future of AI in gaming. It showcases the potential for AI to challenge human limitations and provide new and exciting experiences for gamers.