Vaatenurga võtmist vajavate ülesannete lahendamine stiimulõppe abil
Files
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Võime näha olukorda kellegi teise vaatenurgast on oluline oskus osalemaks mitme agendi koostöös või nendevahelises võistluses. See võime ei ole ainuomane inimestele. Antud töö uurib, mismoodi kellegi teise vaatenurga mudeldamine võiks toimida meie ajus. Selleks loodi virtuaalsetele agentidele keskkond, milles iga agent näeb ainult osa sellest. Näidatakse, kuidas erinevate stiimulõppe meetoditega on võimalik lahendada ülesandeid, mille puhul on kasu teise agendi vaatenurga mudeldamisest. Agendid kasutavad vastase vaatenurga modelleerimiseks tehisnärvivõrke.
Perspective taking is the faculty that allows us to take the point of view of another agent.This capability is not unique to humans. This is an essential ability for agents to achieve efficient social interactions, including cooperation and competition. In this work, we present our progress toward reverse engineering how perspective taking task might be accomplished in our brains. We introduce an environment designed from scratch for the purpose of creating perspective-taking tasks, in which the environment is partially observable by its agents. We also show a set of different models that were able to pass multiple tests that would benefit from perspective taking capabilities. These models weretrained using reinforcement learning algorithms assisted by artificial neural networks.
Perspective taking is the faculty that allows us to take the point of view of another agent.This capability is not unique to humans. This is an essential ability for agents to achieve efficient social interactions, including cooperation and competition. In this work, we present our progress toward reverse engineering how perspective taking task might be accomplished in our brains. We introduce an environment designed from scratch for the purpose of creating perspective-taking tasks, in which the environment is partially observable by its agents. We also show a set of different models that were able to pass multiple tests that would benefit from perspective taking capabilities. These models weretrained using reinforcement learning algorithms assisted by artificial neural networks.