Searching for the interesting stuff in a multi-dimensional parameter space

Abstract

This talk describes work that I have been doing using generative systems and the problems this raises with how to deal with multi-dimensional parameter spaces. In particular I am interested in dealing with problems where there are too many parameters to do a simple exhaustive search, only a small number of parameter combinations are likely to achieve interesting results, but the user still wants to retain creative influence. For a number of years I have been exploring how intricate complex structures may be created by simulating growth processes. In early work, such the Aggregation (Lomas 2005) and Flow series, a small number of parameters controlled various effects that could bias the growth. These could be explored by simply varying all the parameters independently and running simulations to test the results. Simple methods such as these work well when there are up to 3 parameters. However, as the number of parameters increase, the task rapidly becomes increasingly complex, and methods that exhaustively sample all the parameters independently are no longer viable. In this talk I will discuss how I have approached this problem for my recent Cellular Forms (Lomas 2014) and Hybrid Forms (Lomas 2015) works which can have more than 30 parameters, any of which could affect the simulation process in complex and unexpected ways. In particular, systems that have the potential for interesting emergent results often exhibit difficult behavior, where most sets of parameter values create uninteresting regularity or chaos. Only at the transition areas between these states are the most interesting complex results found. To help solve these problems I have been developing a tool called 'Species Explorer'. This uses a hybrid approach that combines both evolutionary and lazy machine learning techniques to assist the user find combinations of parameters that may be worth sampling, helping them to explore for novelty as well as to refine particularly promising results.