Improve agglomerativeClustering for ResponseFitnessClustering

[lisv]

Following the meeting on the 15th of March:

• Perhaps use hashed values instead of strings for the response structure (in ResponseFitnessClustering) to improve performance
• Do not recompute clusters for each generation/individual but base fitness on the probability that a responseObject belongs to one of the existing clusters. Keep the AgglomerativeClustering object small. Recompute clusters after X generations.
• Deal with values that change very often. I.e. if for X many generations, the agglomerative cluster object only contains clusters/individuals with super high dissimarity. Idea: create exclude lists so that this method is no longer used in individuals.
• Issue with random values in response objects: if there is high degree of dissimilarity among individuals with this structure, this method should no longer be used for individuals.
• Implement to not put a value belonging to a key in the feature vector when this key/value pair is also in the request

[lisv]

Update

• Upon further consideration, hashing the structure of the response structure is not necessary since the structures are stored in a HashMap.
• Implemented a second version of clustering that calculates the probabilities of individuals belonging to a certain cluster rather than reclustering every generation. Reclustering is instead done every X generations in this version.
• Key-Value pairs in the response that exist in the request are not put in the feature vector as they make the calculation of distance between responses less precise.