Removing noise with an Autoencoder in a Predator-Prey Ordinary Differential Equation

Sammanfattning: When studying populations of animals, you might want to try to predict the future population growth of the animals within a region. While there are other factors in play and depending on the looked at species, one of the major factors to look at can be the predator-prey relationship between two species. Instead of just looking at the history of the species separately you can create an ordinary differential equation to model the relationship between the two to get a better prediction. Since it is usually impossible to count the entire population within a larger area the collected data will naturally be noisy. This report explores the possibility of removing that noise through the use of an autoencoder. The autoencoder was tested using various methods on a synthetic dataset and then those settings were applied to a real-world dataset. The modified data was then used to create an ODE model to see if the generated model was closer to the original model compared to the one generated by the noisy data. Our results show that the autoencoder performed rather poorly on the synthetic dataset, while it performed well on the real-world dataset. Though the metrics indicated a bad performance on the synthetic dataset, the resulting graphs looked much more accurate to the naked eye which could indicate that the method is at least promising. The autoencoder seems to work better when more noise was added. Therefore, the conclusion is that it is possible with the autoencoder to clean data with heavy noise when constructing an ODE.