Visual Attention in Dynamic Environments and its Application to Playing Online Games

摘要

Abstract In this thesis we present a prototype of Cognitive Programs (CPs) - an executive controller built on top of Selective Tuning (ST) model of attention. CPs enable top-down control of visual system and interaction between the low-level vision and higher-level task demands. 
\n
\nAbstract We implement a subset of CPs for playing online video games in real time using only visual input. Two commercial closed-source games - Canabalt and Robot Unicorn Attack - are used for evaluation. Their simple gameplay and minimal controls put the emphasis on reaction speed and attention over planning. 
\n
\nAbstract Our implementation of Cognitive Programs plays both games at human expert level, which experimentally proves the validity of the concept. Additionally we resolved multiple theoretical and engineering issues, e.g. extending the CPs to dynamic environments, finding suitable data structures for describing the task and information flow within the network and determining the correct timing for each process.