90 likes | 208 Views
On Playing Games without Knowing the Rules. Advisor : Dr. Hsu Presenter : Jia-Hao Yang Author :Denis V. Batalov and B. John Oommen. Outline. Motivation Objective Method Experience Conclusion. Motivation.
E N D
On Playing Games without Knowing the Rules Advisor : Dr. Hsu Presenter : Jia-Hao Yang Author :Denis V. Batalov and B. John Oommen
Outline • Motivation • Objective • Method • Experience • Conclusion
Motivation • We know that one of the interesting areas in AI is to teach machine to play a game against an educated opponent. • But if the machines don’t know the rule of the game?
Objective • This paper will show that the machine will learns the rules of the game, tic-tac-toe, and strategy just as paper’s title say.
Method • To accomplish this goal, we assume that the LM interacts with an environment. • Sense-act-learn procedure • Agent-Environment Interaction Protocol (AEip) • AEip • Because we use JAVA to implement this platform, so we call it JAGUAR
EX Method • AEip specification of Tic-tac-toe • Reinforcement • Doesn’t end the game : -1 • Win & Lose & Tie: + 10 & -10 & +5 • Learning algorithm • Q-learning • Select mathod : If t = 0.1=> P = ∞ (greedy) If t = ∞ => p = 1/j (random)
Experiment • This paper just underscore two set of results • The agents were selecting their actions simultaneously
Experiment • How much faster the agents learn to play when they allowed to make a move on their own turn
Conclusion • In this paper we have specified a novel framework and show how an agent can learn to play a new game without any prior knowledge.