This document proposes using evolutionary algorithms and multi-agent potential fields to develop new artificial intelligence techniques for real-time strategy games. It describes using potential fields to control agent navigation, evolving the parameters through genetic algorithms, and testing the approach in StarCraft scenarios. The results showed the trained potential fields performing comparably to intermediate human players. Further work is suggested in applying this to full RTS games and combining with other optimization techniques.
2. Introduction
• Artificial Intelligence (AI) from RTS games are easy to defeat
• Harder AI are cheating
• Classical solutions like A* and state machines are CPU
intensive.
• “It’s about time” to develop new AI methods
3. Starcraft as test platform
• One of the most popular RTS games.
• You can play three races.
• Possibly the most balanced RTS out there.
4. Starcraft concepts
• Liquipedia definitions:
• Micro: “The ability to control your units individually, in order to
make up for pathing or otherwise imperfect AI.”
• Macro: “The ability to produce units, and keep all of your
production buildings busy.”
• A good player needs to master both techniques.
• An example of good micro (NaDa vultures):
• http://www.youtube.com/watch?v=YXJ5jGCtTYA
5. Potential Fields
• Used for controlling agent
navigation with static and
dynamic obstacles.
• Force fields can be
attractive or repulsive.
• Brighter tiles are more
attractive.
6. Multi-Agent Potential Fields
• Six-step methodology for its design (Hagelbäck & Johansson).
• Thomas Willer Sandberg proposes another step for tuning.
• Seven-step methodology for its design:
• Object identification.
• Potential Fields identification.
• Charge assignation to objects.
• Charge parameters tuning.
• Granurality of time and space assignation.
• Agents of the system identification.
• MAS architecture design.
7. Evolutionary Algorithms (EA)
• Set of parameters = Individuals of the population.
• In each iteration, individuals are recombined and
mutated.
• Better candidates obtain higher fitness function
values.
• The remaining population will be stronger
(Darwin’s natural selection theory).
8. EMAPF-based AI (fields)
• 8 potential fields identified:
• Maximum Shooting Distance attraction.
• Weapon Cool Down repulsion.
• Centroid Of Squad attraction.
• Center Of the Map attraction.
• Map Edge repulsion.
• Own Unit repulsion.
• Enemy Unit repulsion.
• Neutral Unit repulsion.
10. EMAPF-based AI (results)
• 3 Goliaths vs. 6 Zealots:
• http://www.youtube.com/watch?v=VfI8XN91ggU
• Terran Mix vs. Zerg Mix:
• http://www.youtube.com/watch?v=hETcbgybkoc
• 3 Goliaths vs. 20 Zerglings:
• http://www.youtube.com/watch?v=Q0auIScPCYg
11. Conclusions
• It is possible to use EA for tuning potential field parameters.
• Trained potential fields show extraordinary results.
• They are comparable with medium-skilled/advanced players.
12. Future Work
• To use trained potential fields on a Full RTS scenario.
• To develop MAPF-based solutions with different algorithms.
• To study the combination of these techniques with optimization
techniques for macro issues (example: BOs).
• To analyze how difficult is for humans to defeat EMAPF-based AI.