SWORDS-03-10-2016

Slides for my Talk in SWORD in 03/10/2016.

Transcript

1. Prisoners and Spatial Structure Nikoleta Glynatsi SWORDS, October 2016
3. Prisoners and Spatial Structure (3, 3) (0, 5) (5, 0) (1, 1)
4. Strategy 1 class Grumpy(Player): 2 """ 3 A player that gets grumpier the more the opposition defects, 4 and nicer the more they cooperate. 5 Starts off Nice, but becomes grumpy once the grumpiness threshold is hit. 6 Won’t become nice once that grumpy threshold is hit, but must reach a much 7 lower threshold before it becomes nice again. 8 """ 9 10 self.grumpiness = opponent.defections - opponent.cooperations 11 12 if self.state == ’Nice’: 13 if self.grumpiness > self.grumpy_threshold: 14 self.state = ’Grumpy’ 15 return D 16 return C 17 18 if self.state == ’Grumpy’: 19 if self.grumpiness < self.nice_threshold: 20 self.state = ’Nice’ 21 return C 22 return D
5. History Line (1980) (1991) (1992) (2012) (2015)
6. Nowak and May, 1992
7. What do real life interactions look like?
8. What do real life interactions look like? Meta Majority Long Memory Limited Retaliate (0.1/20) AntiCycler Cooperator Shubik Soft Joss: 0.9 Meta Majority Finite Memory ZD-SET-2 Tullock ThueMorse Naive Prober: 0.1 Prober 2 Hard Prober Grofman Punisher Appeaser Eatherley Retaliate (0.1) Grudger Limited Retaliate (0.08/15) Inverse Punisher Inverse DoubleCrosser Hard Go By Majority: 20 Nydegger Cooperator Hunter Davis Firm But Fair Meta Winner Long Memory Bully Anti Tit For Tat Predator Hard Go By Majority: 10 Forgetful Fool Me Once Retaliate (0.05) Nice Average Copier Cycler CCCD Soft Go By Majority: 5 Hard Go By Majority: 40 Forgetful Grudger Hard Go By Majority Meta Winner Finite Memory Cautious QLearner Forgiving Tit For Tat ZD-GTFT-2 Adaptive Win-Stay Lose-Shift Retaliate (0.08) Stochastic WSLS Meta Hunter
9. Measurements 0.0 0.2 0.4 0.6 0.8 1.0 Retaliate (0.05) Forgetful Fool Me Once Soft Go By Majority: 40 Cycler CCCD Anti Tit For Tat Punisher Inverse Win-Shift Lose-Stay ZD-GTFT-2 Adaptive Math Constant Hunter Cycler CCD Hard Go By Majority Win-Stay Lose-Shift SolutionB1 Calculator Eatherley Hard Prober Meta Majority Memory One Tricky Cooperator Grumpy Average Copier Hard Tit For Tat Meta Winner Long Memory Omega TFT Stochastic WSLS Soft Go By Majority: 20 Remorseful Prober: 0.1 Soft Grudger Meta Winner Finite Memory Meta Majority Long Memory Limited Retaliate (0.1/20) Limited Retaliate (0.08/15) Hard Tit For 2 Tats ZD-Extort-4 Cycler DDC Hesitant QLearner Meta Majority Finite Memory Contrite Tit For Tat Naive Prober: 0.1 Davis Meta Winner DoubleCrosser ALLCorALLD Limited Retaliate (0.05/20) Alternator Hunter Champion AntiCycler Adapative Pavlov 2006 Appeaser ZD-GEN-2 Joss: 0.9 ThueMorseInverse Tit For Tat Ripoff Once Bitten Risky QLearner Random: 0.5 Two Tits For Tat Cooperator Grudger Hard Go By Majority: 10 Defector Hunter Alternator Cycler DC Meta Majority Forgetful Grudger ZD-SET-2 Grofman Hard Go By Majority: 20 Cycle Hunter Forgiving Tit For Tat Tricky Defector Nydegger Sneaky Tit For Tat Prober Adapative Pavlov 2011 Cycler CCCCCD Hard Go By Majority: 5 Slow Tit For Two Tats ZD-Extort-2 EvolvedLookerUp Tit For 2 Tats Aggravater Meta Minority BackStabber ThueMorse Shubik Tullock Feld Arrogant QLearner Gradual Firm But Fair Handshake Defector Meta Winner Memory One Prober 2 Inverse Punisher Fool Me Once Fool Me Forever Soft Go By Majority Soft Go By Majority: 5 Cautious QLearner Forgiver Retaliate (0.1) Bully Prober 3 Meta Mixer Hard Go By Majority: 40 Suspicious Tit For Tat SolutionB5 GTFT: 0.33 PSO Gambler Opposite Grudger Meta Hunter ZD-Extort-2 v2 Predator Random Hunter Eventual Cycle Hunter Tester Retaliate (0.08) Fortress4 Cooperator Hunter Thumper Fortress3 Soft Joss: 0.9 Raider e Nice Average Copier Soft Go By Majority: 10 Cyclic Tournament Rankings Inverse Hard Tit For 2 Tats Meta Majority Long Memory Nydegger Thumper Ripoff Prober 3 Math Constant Hunter Grofman Fortress3 Eatherley Defector Naive Prober: 0.1 ThueMorse Contrite Tit For Tat Fool Me Forever ThueMorseInverse Tricky Cooperator Meta Majority Random Hunter Meta Mixer Stochastic WSLS Cooperator Hard Tit For Tat Predator Hard Go By Majority Arrogant QLearner Calculator Aggravater Random: 0.5 Fortress4 DoubleCrosser Cooperator Hunter Cycle Hunter Soft Go By Majority: 10 Retaliate (0.05) Grudger Hard Go By Majority: 10 Tester SolutionB1 Feld Limited Retaliate (0.1/20) Soft Go By Majority: 5 ZD-GEN-2 Retaliate (0.1) Alternator Eventual Cycle Hunter Anti Tit For Tat Forgiving Tit For Tat Adaptive Hesitant QLearner Hard Go By Majority: 20 Sneaky Tit For Tat Prober e Soft Grudger ZD-Extort-2 v2 Davis Win-Shift Lose-Stay ALLCorALLD Defector Hunter AntiCycler Prober 2 Suspicious Tit For Tat Gradual ZD-Extort-4 Handshake Average Copier Cycler CCD Soft Go By Majority: 20 ZD-GTFT-2 Soft Go By Majority Tricky Defector Risky QLearner Nice Average Copier Bully Cycler CCCD Adapative Pavlov 2011 Cycler CCCCCD Hard Go By Majority: 5 ZD-Extort-2 Firm But Fair Meta Hunter Two Tits For Tat Punisher Alternator Hunter Forgetful Fool Me Once GTFT: 0.33 Forgiver Meta Majority Finite Memory Opposite Grudger Hard Go By Majority: 40 Limited Retaliate (0.08/15) Win-Stay Lose-Shift Tullock Once Bitten Limited Retaliate (0.05/20) Retaliate (0.08) Hard Prober Tit For 2 Tats Soft Go By Majority: 40 Cycler DDC PSO Gambler BackStabber Meta Majority Memory One Shubik Meta Winner Finite Memory ZD-SET-2 Cycler DC Meta Winner Meta Minority Appeaser Soft Joss: 0.9 Tit For Tat SolutionB5 Joss: 0.9 Meta Winner Memory One Champion EvolvedLookerUp Omega TFT Meta Winner Long Memory Slow Tit For Two Tats Cautious QLearner Grumpy Forgetful Grudger Remorseful Prober: 0.1 Inverse Punisher Raider Fool Me Once Adapative Pavlov 2006 Periodic Lattice Tournament Rankings Player Rankings by Winning Ration in Cyclic vs Periodic Lattice Topologies size 50
10. Training a Strategy using Genetic Algorithm
11. Parents: Crossovers: Children: Mutation: ...CDCCDDCDDCCCDDCC... ...CCDDCDCDCDCDCCDC... ...DDDDDCDDCDCCCDCD... ...CCCCDCDDDCDDDDDC... ...CDCCDDCDDCCCDcrossoverDCC... ...CCDcrossoverDCDCDCDCDCCDC... ...CDCCDDCDDCCCDCCD... ...DCCDCDCDCDCDCCDC... ...CDCCDDCCDCCCDCCD... ...DCCDCDCDCDCDDCDC...
12. Conclusions and Futher Research Conclusions: – The topology affects the strategies performance – Using regression we can predict 2/132 strategies behaviour – None of the 132 strategies performed well in all experiments – For specific spatial tournaments a satisfactory strategy has been trained
13. Conclusions and Futher Research Plans for PhD: – Game Theory – Machine Learning
14. @NikoletaGlyn https://github.com/Nikoleta-v3 https://github.com/Axelrod-Python/Axelrod