Скачать или смотреть SAPEO for Games [CoG 19 Short Video Competition]

Optimisation problems, such as posed in search-based procedural content generation, have several characteristics that make them difficult to handle. The landscape of the fitness function is often unknown and not differentiable, and contains several non-explicit constraints. In addition, many fitness values are based on AI playthroughs, making them expensive to evaluate as well as noisy.

Many optimisation tasks in engineering and other real-world applications have similar characteristics. These problems are usually approached with surrogate-assisted evolutionary algorithms. However, most of these algorithms assume that the surrogate model is able to make good predictions for fitness values after a reasonable number of samples. However, this is likely not true in many complex applications, such as games. The SAPEO algorithm (Surrogate-Assisted Evolutionary Optimisation) was developed to mitigate this issue by controlling the risk of errors introduced from model prediction errors.

In this video, I explain how SAPEO works and why it is suitable for application to optimisation problems in games.

-----------------------------------
Related publication: Vanessa Volz. "Uncertainty handling in surrogate assisted optimisation of games", PhD Thesis, TU Dortmund University. 2019. http://dx.doi.org/10.17877/DE290R-19924

Code Github: https://github.com/TheHedgeify/uncert...
Code CodeOcean: https://codeocean.com/capsule/2600355

-----------------------------------
Created with Animaker
Submission for CoG 19 Short Video Competition