Desenvolvimento de Inteligências Artificiais Baseadas em Planeamento de Monte Carlo Temporal para Videojogos de Ação Furtiva

Development of Artificial Intelligence Systems for Stealth Games based on the Monte Carlo Method FEUP - 2013/2014 Presented by: Diogo Silva Supervisors : Prof. Eugénio Oliveira Msc. Pedro Nogueira AI in Video Games Starcraft - Blizzard(1998) 1 Introduction - Context Spore - Maxis(2008)Empire Total War - The Creative Assembly (2009) Motivation Phyllimar - DeviantArt (2012) 2 Introduction - Motivation Objectives 3 Introduction - Objectives ● Dynamic action planning ● Stealth movement and behaviour ● Creation of a testing platform Intelligent Agents in Video Games 4 State of the Art - Agents in Video Games ● Autonomy ● Reaction ● Proactiveness Environment Agent Agent Architectures 5 State of the Art - Architectures ● Deliberative Architectures ○ Depend on the representation of knowledge ○ Symbolic reasoning ○ Belief-Desire-Intention(BDI) Agent Architectures 6 State of the Art - Architectures ● Reactive Architectures ○ No internal state, only reactions ○ Instant feedback ○ Subsumption Architectures Agent Architectures 7 State of the Art - Architectures ● Hybrid Architectures ○ Contain parts of previous architectures ○ Layers communicate between them ○ 3-Tiered Architecture Current Agents 8 State of the Art - Agents FSMs by Fernando Bevilacqua Agent Implementation 15 Implementation - Agent ● Sensors ● Knowledge ● Goals ● Actions ● Planning ● Movement Agente Sensors 16 Implementation - Sensors Sight Knowledge Representation 17 Implementation - Knowledge ● Map Knowledge ○ Known Actions ○ Hiding Places ○ Known map ○ Danger spots ● Agent State ○ Position ○ Inventory ○ Hitpoints Goal-Oriented Action Planning 18 Implementation - GOAP Goals Explore Hide Loot Escape Actions Hide Look Move Steal Peek Planning 19 Implementation - Planning Planning 20 Implementation - Planning Monte-Carlo Tree Search 21 Implementation - MCTS Stealth Movement 22 Implementation - Movement Testing Platform 23 Implementation - Testing Platform ● Parameters ○ Map size ○ Room size ○ Maximum room number ● Stealth Elements ○ Furniture ○ Guards ○ Lights/Shadows Simulations 24 Results - Simulations ● Random: Random Agent ● MCTS-F: Forward Search ● MCTS-B: Backward Search Objective fulfillment 31 Conclusions - Objective fulfillment ● Dynamic action planning ● Stealth movement and behaviour ● Creation of a testing platform Conclusions 32 Conclusions ● Forward-Search: Safer choice for video games ● Backward-Search: More effective choice depending on time ● Possibility of usage in serious games ○ Security Training ○ Stealth behaviour in animals Future Work 33 Conclusions - Future Work ● Reduce knowledge representation complexity ● Personality Parameters ● Upgrade the testing platform ● Create new Actions/Goals for the Agent Thank you Eric Joyner 2007 34