Understanding Player Behavior in Plaicraft.ai for AGI Research by Aamna, Oliver, Sanchia, and Stephanie

Introduction

The Pacific Laboratory for Artificial Intelligence (PLAI) at UBC is conducting groundbreaking research on human behavior in video game environments using generative AI and machine learning. A centerpiece of this work is Plaicraft.ai, a public platform allowing players to engage with Minecraft in a cloud environment to collect valuable gameplay data. PLAI's ultimate goal is to collect over 10,000 hours of gameplay to advance research in artificial general intelligence (AGI).

Research Objective

PLAI researchers aim to identify which types of players are most likely to contribute significant gameplay data. This understanding will:

• Optimize recruitment strategies to target specific player profiles.
• Improve resource allocation for software and infrastructure.

To address this objective, we analyzed how player characteristics predict total gameplay hours (played_hours) by building and evaluating three models:

1. K-NN Regression: Predicting played_hours using age.
2. Simple Linear Regression: Predicting played_hours using age.
3. Multivariable Linear Regression: Predicting played_hours using age, subscription status, experience level, and number of sessions.

We hypothesized that the multivariable linear regression model would outperform the other models, with lower age, high experience level, positive subscription status, and a high number of sessions correlating with higher played_hours.

---

Summary of Findings

Model Performance

• Multivariable Linear Regression achieved the lowest root mean squared prediction error (RMSPE) of 6.59, outperforming the other models.
• K-NN Regression and Simple Linear Regression had much higher RMSPE values of 29.17 and 30.00, respectively.

These results indicate that multivariable linear regression was the most effective model, likely due to its ability to account for multiple predictors.

Key Insights

1. Player Age and Gameplay Hours

   • Younger players tend to log more gameplay hours, aligning with the expected trend that teens and young adults (aged 15-25) are more engaged with gaming.
   • However, the dataset was heavily skewed towards younger players, which may limit the model's predictive performance for older age groups.

2. Variable Relationships

   • Age alone did not show a clear linear relationship with played_hours, making simple linear regression less effective.
   • Number of sessions had a stronger linear relationship with played_hours, highlighting its importance as a predictor.

Unexpected Findings

• K-NN Regression performed only slightly better than simple linear regression. This may be due to limited data variety in the age range, with most participants being under 25 years old.

---

Challenges and Future Improvements

• Data Imbalance: The dataset lacked diversity in player age, reducing the generalizability of our findings. Future efforts should focus on recruiting players from a broader age range.
• Model Refinement: Exploring additional predictive algorithms (e.g., random forests or gradient boosting) may improve accuracy.
• Feature Engineering: Incorporating more behavioral metrics, such as in-game actions or achievements, could enhance model performance.