Towards Automatic StarCraft Strategy Generation Using Genetic Programming
1 like4,581 views
The document outlines a framework for automatically generating non-human strategies in Starcraft using genetic programming, aiming to explore new game plans. It details the experimental setup, including two fitness functions (victory-based and report-based) and the results showing that the generated strategies can outperform human-coded ones. Future work includes further evaluations and potential incorporation of machine learning techniques.
Towards Automatic StarCraft Strategy Generation Using Genetic Programming
- 1. Towards Automatic StarCraft Strategy Generation Using Genetic Programming Pablo García Sánchez, Alberto Tonda, Antonio M. Mora, Giovanni Squillero, J.J. Merelo University of Granada, Spain Politecnico di Torino, Italy INRA, France
- 2. Objective Automatic generation of non- human strategies for RTS, to explore new possible game plans.
- 3. Objective Automatic generation of non- human strategies for RTS, to explore new possible game plans. Also, it’s F***ING COOL!
- 4. Outline ● Introduction ● Framework Description ● Experimental Setting ● Results ● Conclusions and Future work
- 5. Introduction ● Starcraft: de-facto testbed for RTS research ● Motivation: generate high-level strategies using Genetic Programming ● Two different fitness functions used: victory- based and report-based. ● Validation against bots not used during evaluation.
- 9. Framework description: fitness ● Victory-based: lexicographic victory counts using the default final score returned by StarCraft at the end of one match against each of 12 different opponents (divided in 3 tiers). ● Report-based: a more complex metric, aiming at separating military success from in-game economy development, using different rewards, computed after 3 matches against each of 4 different opponents.
- 10. Experimental Setting: Parameters Parameter Value Meaning μ 30 Population size λ 30 Number of genetic operators applied at each generation σ 0.9 Initial strength of the mutation operators α 0.9 Inertia of the self-adapting process τ (1,4) Size of the tournament selection MG 30 Stop condition: maximum number of generations
- 11. Experimental setting: parrying bots ● Race evolved: Zerg ● Bot used as base to evolve the high level strategy: OpprimoBot ● Victory-based fitness hand-coded enemies: ○ Tier 1 (Easy): TerranDummy, ProtossReaverDrop1B, ProtossDefensive, ZergHydraMuta ○ Tier 2 (Medium): OBProtossTemplarRush, ProtossReaverDrop, TerranDefensiveFB, TerranDefensive ○ Tier 3 (Hard): ZergLurkerRush, TerranWraithHarass, TerranPush, TerranMarineRush ● Report-based fitness hand-coded enemies: ○ TerranDummy, ProtossReaverDrop, TerranWraithHarass, ZergLurkerRush (repeating 3 times against each strategy).
- 12. Results: best individuals Best individual Average of population Tier 3 4 1.73 Tier 2 3 1.83 Tier 1 3 2.93 Score ratio 0.0481 0.0378 Best individual Average of population Military Victories 3 1.76 Economic victories 1 1.93 Relative destruction 400245 358172 Time to loss 1120 1380.7 Relative Economy 0.309 0.501 Victory-based fitness Report-based fitness
- 13. Results: Validation of best bots Vs. Bot Victory-based Report-based OBTerranDefensiveFB 7 1 OBProtossTemplarRush 4 8 OBZergHydraMuta 10 1 OBZergLurkerRush 8 0 OBProtossDefensive 8 5 OBProtossReaverDrop1B 5 1 OBTerranDefensive 5 1 OBProtossReaverDrop 3 6 OBTerranMarineRush 7 0 OBTerranWraithHarass 5 0 OBTerranPush 6 3 OBTerranDummy 10 10 Victory-Based * 8 Report-Based 2 * OpprimoBot 6 1 TOTAL 86 of 140 45 of 140
- 14. (Open) Challenges ● Evaluation time ○ Individuals compared vs many bots on many maps ○ Two virtual machines per game (!) ○ Even with limitless resources, 10’ per generation ● Rules and actions ○ Level of abstraction (meta-strategy, strategy, tactic) ○ Constraining the GP, or giving it more freedom? ● Bot generality ○ How to avoid overfitting on one map/vs one race? ○ Many opponents! But evaluation time increases
- 15. Conclusions ● Our framework has been able to generate high-level strategies that outperform (naive) human-coded ones ● Victory-based fitness outperforms Report- based fitness ● Future work: ○ more evaluations, maps, races (Protoss and Terrans), co-evolution, machine learning, map analysis.
- 16. Thank you!