Embedded Files
Plot 1: Our recorded data and coherences plotted against known scores for those play-throughs. While it is not accurate in terms of specific scores, the general trend is reflective of a 0-5 star ranking system.
Score Prediction
Score Prediction
We can predict your Just Dance score to within 2000 points. While not perfect accurate, that level of precision means we can accurately rank your play-through on the game's range of 0-5 stars (⭐).
In order to do so, we make use of our cohere function and our calculated linear regression model. By finding the coherence of a given play-through and passing it into our defined linear model, we output a score and ranking.
Running our regression for the two trial runs shown to the left returns values of 3 and 2 respectively. Given a score of 7040, Just Dance officially gives a ranking of 3 ⭐s, which aligns exactly with what we predict. The same is true for the second example.
Code 1: A snippet of code showing how we evaluated our defined linear polynomial model. We divide the result of the regression by 2000 and then floor the result to get the closest ⭐ ranking.
Reflection
Reflection
Limitations
Limitations
Next Steps
Next Steps
- Due to our method of analysis, our regression is restricted to a single song
- We are analyzing the acceleration data over the course of the entire song rather than move-by-move, which reduces our prediction resolution.
- We are utilizing the magnitude-squared coherence between two signals, which does not take into account the time domain. Because of this, a signal with equivalent frequencies but which is out of sink to the game will still yield a high coherence and predict a high score, despite being low in reality.
- Our regression model is trained on a biased dataset, which means that our regression is more likely to make predictions within the 6-8k range rather than uniformly across the 0-11k range.
Our current model is not as accurate as we would like. However, we've identified a few things that we hope could increase the reliability of our model:
- We want to collect more data for low-scoring runs. We think that including a more uniform distribution of data from across the score spectrum will increase the robustness of our regression model.
- We want to explore simple signal filtering. Due to the nature of our motions, our accelerometer data is very complex and noisy. By filtering out unwanted frequencies and smoothing out desired ones, we hope to increase our model's reliability.
- Through the use of more advanced machine learning algorithms like convolutional neural networks, we could perform more robust pattern detections.
- Expand to more songs in Just Dance
- Explore other methods of processing signals to determine score
- Collect accelerometer data from Wii Remote itself
Page updated
Google Sites
Report abuse