43
Adaptive Reinforcement Learning Agents in RTS Games Eric Kok

Adaptive Reinforcement Learning Agents in RTS Games

  • Upload
    elvina

  • View
    43

  • Download
    6

Embed Size (px)

DESCRIPTION

Adaptive Reinforcement Learning Agents in RTS Games. Eric Kok. 1.Introduction. Abstract. Motivation. RTS Games. Bos Wars. Agent Technology. Reinforcement Learning. 2 .Existing Research. Dynamic Scripting. Evolutionary Learning. Case-based Plan selection. - PowerPoint PPT Presentation

Citation preview

Adaptive Reinforcement Learning Agents in RTS Games

Adaptive Reinforcement Learning Agents in RTS GamesEric Kok1.IntroductionAbstract

3Motivation

RTS Games

Bos Wars

Agent Technology

No need to explicitly state how to solve the problemBetter reusable7Reinforcement Learning

To work in an UnKnown EnviromentTypes of RLExploration-exploitation problem82.Existing Research9Dynamic Scripting

bound to episodic tasks, has a fixed script size, uses once-only rule selection and can't reuse possibly mutually effective rules between situations. 10Evolutionary Learning

Offline Learning11Case-based Plan selectionNo comment !12Concurrent hierarchal Reinforcement LearningReduce the search space133.Learning winning RTS Game Strategies14Learning in 2apl agents15BDI Agents16Temporal Difference LearningEligibility Traces must be implemented17Monte-Carlo Methods184.Improvements implemented19Rule Guards20Softmax Exploration Policy21Strategy Hierarchy22MC-DS Hybrid235.Improvements out of project scope24Dynamic Scripting Enhancement25Comparison of more Learning algorithms26Game State Function Approximation27Semi-MDP Reinforcement LearningConsidering time286.Adaptation to opponent Strategies29Implicit AdaptationIncluding opponents data30Explicit Adaptation31Multi-Agent Learning328.Adaption Ideas Out of scope33Explicit Adaptation through player modeling34Opponent Strength AdaptationMore fun359.Project ConclusionsMore fun36Fail Rates

More fun37Av. Turning Points

More fun38Usability of Learning agents in computer gamesMore fun3910.Future ResearchMore fun40Learning on a full complex game taskMore fun41Strategy Visualization ToolMore fun42Explicit Player ModelingMore fun43