Adaptive Game AI improves adaptability ofopponent AI as well as the challenge level of thegameplay; as a result the entertainment of game isaugmented. Opponent game AI is usually implementedby scripted rules in video games, but the most updatedalgorithm of UCT (Upper Confidence bound for Trees)which perform excellent in computer go can also beused to achieve excellent result to control non-playercharacters (NPCs) in video games. In this paper, theprey and predator game genre of Dead End is used asa test-bed, the basic principle of UCT is presented, andthe effectiveness of its application to game AIdevelopment is demonstrated. The experimentcompares the performance of different NPCs controlapproaches: given a 300 milliseconds for eachsimulation step, the approach of UCT with recognizedplayer’s strategy pattern is better than the one of UCTwithout recognized player’s strategy pattern, the worstone is Monte-Carlo approach.