Applying fuzzy control in fighting game ai
- 1. Applying Fuzzy Control in Fighting Game AI Chun Yin Chu and Ruck Thawonmas Intelligent Computer Entertainment Lab. Graduate School of Information Science and Engineering Ritsumeikan University
- 2. In short… 1: Kaito Yamamoto, Syunsuke Mizuno, Chun Yin Chu and Ruck Thawonmas, "Deduction of Fighting-Game Countermeasures Using the k-Nearest Neighbor Algorithm and a Game Simulator," Proc. of 2014 IEEE Conference on Computational Intelligence and Games (CIG 2014), Dortmund, Germany, pp. 437–441, Aug. 26-29, 2014. • Applying fuzzy control in fighting game AI is able to alleviate the “cold start” problem of k-nn prediction technique1 • Fuzzy control is found to be more effective than crispy rules
- 3. What is Fighting Game? Board Game Turn-based, slow pace Well researched Fighting Game Real time, fast pace Emergent research field
- 4. FightingICE • First fighting game AI platform for annual competitions • Written in Java, suitable for AI development by small team • 3 rounds in a game • 60 frame/second • Data delayed by 15 frames • Imperfect information • Previously 2 categories • 1 Character • 3 Characters • Currently 1 character http://www.ice.ci.ritsumei.ac.jp/~ftgaic/index.htm
- 5. Existing techniques: k-nn Prediction & Simulation1 • AI is re-trained for each game • Use a game simulator to predict the current game state to counter the delay in information • Use k-nn for predicting the opponent’s next attack action (skills) among 31 possible candidates • Use another game simulator to decide a counter-measure action among a set of major actions 1: Kaito Yamamoto, Syunsuke Mizuno, Chun Yin Chu and Ruck Thawonmas, "Deduction of Fighting-Game Countermeasures Using the k-Nearest Neighbor Algorithm and a Game Simulator," Proc. of 2014 IEEE Conference on Computational Intelligence and Games (CIG 2014), Dortmund, Germany, pp. 437–441, Aug. 26-29, 2014.
- 6. Performance of k-nn Prediction & Simulation AI (MizunoAI) • MizunoAI is able to defeat the champion of the 2013 competition • Score > 500 is considered victory in a round Score of MizunoAI vs T, the 2013 champion
- 7. Cold Start • MizunoAI is re-trained for each game • If there is not enough opponent action data (=at the beginning of the game)… • No data to predict opponent’s action! The AI just keeps guarding! • MizunoAI is thus weak during the early stage of the fight => cold start Cannot predict and simulate!
- 8. Applying Fuzzy Control Is the number of opponent data sufficient? sufficient Use kNN prediction and simulation kNN prediction simulation …… Use rule-base. Distance? insufficient close far Use close- range attack Use long-range attack Fuzzy Fuzzy
- 9. Fuzzy Control – Use k-nn prediction? Singleton Defuzzification: Probability(“use_MizunoAI”) = sufficient(oppData) e.g. if oppData == 10, the probability of MizunoAI being used is 0.5 0 0.25 0.5 0.75 1 0 5 10 15 20 Membership oppData (number of opponent data) Membership function for "sufficient data" and "insufficient data" insufficient sufficient RULE 1 : IF oppData IS sufficient THEN decision IS use_knn; RULE 2 : IF oppData IS insufficient THEN decision IS rule_base; oppData is the number of opponent action record in the corresponding dataset
- 10. Fuzzy Control – close or far? Only the horizontal distance of the two characters is considered Singleton Defuzzification: Probability(“close-range-attack”) = close-range(distance); e.g. if distance == 60, the probability of using close-range attack is 0.5 RULE 1 : IF distance IS close THEN decision IS close_range_attack; RULE 2 : IF distance IS far THEN decision IS long_range_attack; 0 0.2 0.4 0.6 0.8 1 0 40 80 120 160 200 240 280 320 360 Membership distance Membership function for “close distance" and “far distance” close-range long-range
- 11. Evaluation • 200 3-round games (100 games played as P1 + 100 games as P2) against: • MizunoAI: original k-nn prediction and simulation AI as proposed by Yamamoto et al. (2014) • ChuMizunoAI_crispy: AI proposed by this paper, but replaced fuzzy control with crispy threshold • All AIs used 3 as the k-value for the k-nn algorithm • All AIs used the character KFM and fought on FightingICE v.1.02 Round 1 Round 2 Round 3 Average 542 502 514 SD 104 125 128 Results 129 wins 3 draws 68 loses 67 wins 70 draws 63 loses 83 wins 52 draws 65 loses VS ChuMizunoAI_crispy Threshold: oppData = 10, distance = 60 Round 1 Round 2 Round 3 Average 647 572 542 SD 129 166 146 Results 174 wins 26 loses 71 wins 110 draws 19 loses 56 wins 119 draws 25 loses VS MizunoAI • Better performance than the original MizunoAI • Fuzzy control fared better than crispy rule MizunoAI performed poorly due to cold start Fuzzy rule-base outperforms crispy rules
- 12. Video: ChuMizunoAI vs MizunoAI P1: ChuMizunoAI P2: MizunoAI
- 13. Fighting Game AI Competition 2015 Fighting Game AI Competition will be held at CIG 2015 Midterm Deadline: 7 June 2015 Final Deadline: 16 August 2015 Please refer to http://www.ice.ci.ritsumei.ac.jp/~ftgaic/
- 14. Questions are welcome Thank you for your kind attention
- 15. Appendix 1 – Fuzzy Rule-Based Move Sets STAND_B STAND_D_DB_BA THROW_A close, to-ground STAND_D_DF_FC/A STAND_D_DB_BB AIR_D_DB_BB far, to-ground STAND_F_D_DFA CROUCH_FB STAND_F_D_DFB close, to-air STAND_D_DF_FB AIR_D_DF_FA AIR_D_DB_BB far, to-air to-air to-ground close far
- 16. Appendix 2 – On Singleton Defuzzification • All fuzzy controls use singleton defuzzication • The membership values are directly used as probabilities, which is then used to derive a single defuzzified output 0 0.25 0.5 0.75 1 0 5 10 15 20 Membership oppData Membership function for "sufficient data" and "insufficient data" insufficient sufficient oppData = 6 P(“use_mizunoAI”) : P(“rule_base”) = 1 : 9 oppData = 13 P(“use_mizunoAI”) : P(“rule_base”) = 8 : 2
- 17. Appendix 3 – Evaluation using ChuMizunoAI_crispy with different parameters Round 1 Round 2 Round 3 Average 518 517 510 SD 57 47 62 Results 12 wins 1 draw 7 loses 8 wins 7 draws 5 loses 7 wins 6 draws 7 loses Threshold: oppData = 5, distance = 40 • 20 3-round games for each set of parameters (10 as P1, 10 as P2) • All against ChuMizunoAI_crispy • All figures are corrected to nearest integer Round 1 Round 2 Round 3 Average 549 507 501 SD 90 78 97 Results 14 wins 1 draw 5 loses 8 wins 5 draws 7 loses 9 wins 5 draws 6 loses Threshold: oppData = 5, distance = 60 Round 1 Round 2 Round 3 Average 542 523 544 SD 101 93 104 Results 14 wins 6 loses 8 wins 5 draws 7 loses 9 wins 6 draws 5 loses Threshold: oppData = 5, distance = 80
- 18. Appendix 3 – Evaluation using ChuMizunoAI_crispy with different parameters Round 1 Round 2 Round 3 Average 540 517 518 SD 78 64 78 Results 14 wins 3 draws 3 loses 6 wins 10 draws 4 loses 8 wins 10 draws 2 loses Threshold: oppData = 10, distance = 40 Round 1 Round 2 Round 3 Average 546 506 511 SD 97 60 85 Results 12 wins 8 loses 8 wins 7 draws 5 loses 8 wins 6 draws 6 loses Threshold: oppData = 10, distance = 80
- 19. Appendix 3 – Evaluation using ChuMizunoAI_crispy with different parameters Round 1 Round 2 Round 3 Average 517 525 502 SD 84 39 90 Results 12 wins 1 draws 7 loses 8 wins 10 draws 2 loses 7 wins 8 draws 5 loses Threshold: oppData = 15, distance = 40 Round 1 Round 2 Round 3 Average 501 507 504 SD 56 70 96 Results 11 wins 9 loses 7 wins 9 draws 4 loses 7 wins 9 draws 4 loses Threshold: oppData = 15, distance = 60 Round 1 Round 2 Round 3 Average 517 516 528 SD 99 62 123 Results 13 wins 7 loses 11 wins 4 draws 5 loses 8 wins 6 draws 6 loses Threshold: oppData = 15, distance = 80