Towards an Agent-Based Independent Component Analysis 
Roman Belavkin
Abstract 
A question of using decision-theoretic agents for Independent Components Analysis (ICA) is 
investigated. In ICA, the observable signal is a mixture of statistically independent 
sources. ICA algorithms find transormations of the observable mixture that minimise the 
dependence between its components (e.g. by minimising mutual information or maximising the 
entropy or its approximations for the observable). This triple (the observable, the measure 
of independence and the transformations) is used as the input, utility and actions of a 
decision-theoretic agent. Using Bayesean inference on the Markov transition model between 
input states (observables), utilities (their independce) and actions (tranformations), the 
agent learns which transformations decompose the input signal into a set of independent 
sources. In addition, the question of using communities of agents is also discussed. For 
the linear ICA, the output of an agent can be the demixing matrix, but the conpcept could 
potentially be applied to the non-linear ICA.