A Connectionist Approach to Spatial Memory and Planning

Abstract

This c hapter describes the design and testing of a vision?based model of spatial memory ?SM?? Firstly ? three theories of biological SM are discussed? The Stim ulus? Reaction theory ?e?g? Hull ???????? the cognitiv e map theory of T olman ?????? and the T opological Net ork?map theory of Byrne ??????? The w third one is the most promising theory of SM in h umans and poten tially allo ws for the most ?exible planning at the lo w est computational and memory cost in robot implemen tations? Secondly ? the T opological Net ork?map theory is translated in w to general principles? i? successiv e perceptual states are link ed? ii? pairs of states are associated with the action responsible for the transition bet een them iii? links are exploited for building goal?directed sequences w of states during planning and iv? bac kw ard planning is preferred to forw ard planning? Tw o forms of connectionist implemen tation of these principles are discussed? using a sparse or a distributed represen tation of states? The sparse approac h is the simplest and leads to a SM in the form of a view?graph ?Sc olk h? opf and Mallot? ?????? The sparse neural net ork w model described here is c haracterized b y? i? the use of a resistiv e?grid paradigm for bac kw ard planning? transforming the view?graph in to a dynamic v alue?map? ii? the use of a separate net ork represen w ting an in erse model of the robot?en v vironmen tin teraction? and iii? the use of state?object associations to driv e planning b y object?fetc hing tasks? Thirdly ? planning and map?learning experimen ts are performed with a robot? The perceptual states are de?ned as images from a narro w? angle on?board video camera? Objects are represen ted b y letters dra wn on cards placed in the en vironmen t? The instruction ?fetc hobjectA?is translated in to ?locate letter A?? These experimen ts rev eal three problems with the implemen tation of the view?graph principle? i? local?minima due to in terference bet een w disjoin t sequences of views built during exploration? ii? ?represen tational dead?ends? in the form of views without a link to other views and iii? aliasing? The causes of these problems are discussed and solutions proposed? Solutions include a more complex vision system and a mec hanism for in ternal in tegration of in termediate steps of plans.