Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
UCT for tactical assault planningRadha-Krishna Balla and Alan Fern, 2009
Seminar paper
Đorđe Relić
University of Basel
November 2015
Content overview
Problem: tactical planning in RTS games
AI planners for RTS games
Monte Carlo and UCT
UCT in tactical assault planning
Experimental results
November 2015 Đorđe Relić @ University of Basel 2
Real-time strategy games
November 2015 Đorđe Relić @ University of Basel 3
Warcraft II (1995 by Blizzard)
Problem : defeat the enemy
Three sub-problems:
– Unit control
– Strategic planning
– Tactical planning
November 2015 Đorđe Relić @ University of Basel 4
Real-time strategy games
AI planners for RTS games
Choosing an action sequence that leads to victory
– Scripting
– Dynamic scripting
– Monte Carlo planning
November 2015 Đorđe Relić @ University of Basel 5
Photo credits Microsoft Research
November 2015 Đorđe Relić @ University of Basel 6
Multi-armed bandit
UCT
November 2015 Đorđe Relić @ University of Basel 15
UCB rule: 𝑄+(𝑠, 𝑎) = 𝑄 𝑠, 𝑎 + 𝑐 ×log 𝑛 𝑠
𝑛 𝑠 , 𝑎
𝑛 𝑠, 𝑎 ← 𝑛 𝑠, 𝑎 + 1 ; 𝑛 𝑠 ← 𝑛 𝑠 + 1
𝑄 𝑠, 𝑎 ← 𝑄 𝑠, 𝑎 +1
𝑛 𝑠, 𝑎[𝑅 − 𝑄 𝑠, 𝑎 ]
Experiment results
Wargus and Stratagus
12 scenarios
Two objective functions
Comparison with baseline algorithms:– Random
– Attack closest
– Attack weakest
– Stratagus AI
– Human player
November 2015 Đorđe Relić @ University of Basel 17
Conclusion
Brief intro to RTS games
UCT algorithm
UCT applied to RTS tactical assault
Experimental results
November 2015 Đorđe Relić @ University of Basel 22