๐Ÿ˜‡
Deep Multi-Agent Reinforcement Learning
  • Deep Multi-Agent Reinforcement Learning
  • Abstract & Contents
    • Abstract
  • 1. Introduction
    • 1. INTRODUCTION
      • 1.1 The Industrial Revolution, Cognition, and Computers
      • 1.2 Deep Multi-Agent Reinforcement-Learning
      • 1.3 Overall Structure
  • 2. Background
    • 2. BACKGROUND
      • 2.1 Reinforcement Learning
      • 2.2 Multi-Agent Settings
      • 2.3 Centralized vs Decentralized Control
      • 2.4 Cooperative, Zero-sum, and General-Sum
      • 2.5 Partial Observability
      • 2.6 Centralized Training, Decentralized Execution
      • 2.7 Value Functions
      • 2.8 Nash Equilibria
      • 2.9 Deep Learning for MARL
      • 2.10 Q-Learning and DQN
      • 2.11 Reinforce and Actor-Critic
  • I Learning to Collaborate
    • 3. Counterfactual Multi-Agent Policy Gradients
      • 3.1 Introduction
      • 3.2 Related Work
      • 3.3 Multi-Agent StarCraft Micromanagement
      • 3.4 Methods
        • 3.4.1 Independent Actor-Critic
        • 3.4.2 Counterfactual Multi-Agent Policy Gradients
        • 3.4.2.1 baseline lemma
        • 3.4.2.2 COMA Algorithm
      • 3.5 Results
      • 3.6 Conclusions & Future Work
    • 4 Multi-Agent Common Knowledge Reinforcement Learning
      • 4.1 Introduction
      • 4.2 Related Work
      • 4.3 Dec-POMDP and Features
      • 4.4 Common Knowledge
      • 4.5 Multi-Agent Common Knowledge Reinforcement Learning
      • 4.6 Pairwise MACKRL
      • 4.7 Experiments and Results
      • 4.8 Conclusion & Future Work
    • 5 Stabilizing Experience Replay
      • 5.1 Introduction
      • 5.2 Related Work
      • 5.3 Methods
        • 5.3.1 Multi-Agent Importance Sampling
        • 5.3.2 Multi-Agent Fingerprints
      • 5.4 Experiments
        • 5.4.1 Architecture
      • 5.5 Results
        • 5.5.1 Importance Sampling
        • 5.5.2 Fingerprints
        • 5.5.3 Informative Trajectories
      • 5.6 Conclusion & Future Work
  • II Learning to Communicate
    • 6. Learning to Communicate with Deep Multi-Agent ReinforcementLearning
      • 6.1 Introduction
      • 6.2 Related Work
      • 6.3 Setting
      • 6.4 Methods
        • 6.4.1 Reinforced Inter-Agent Learning
        • 6.4.2 Differentiable Inter-Agent Learning
      • 6.5 DIAL Details
      • 6.6 Experiments
        • 6.6.1 Model Architecture
        • 6.6.2 Switch Riddle
        • 6.6.3 MNIST Games
        • 6.6.4 Effect of Channel Noise
      • 6.7 Conclusion & Future Work
    • 7. Bayesian Action Decoder
      • 7.1 Introduction
      • 7.2 Setting
      • 7.3 Method
        • 7.3.1 Public belief
        • 7.3.2 Public Belief MDP
        • 7.3.3 Sampling Deterministic Partial Policies
        • 7.3.4 Factorized Belief Updates
        • 7.3.5 Self-Consistent Beliefs
      • 7.4 Experiments and Results
        • 7.4.1 Matrix Game
        • 7.4.2 Hanabi
        • 7.4.3 Observations and Actions
        • 7.4.4 Beliefs in Hanabi
        • 7.4.5 Architecture Details for Baselines and Method
        • 7.4.6 Hyperparamters
        • 7.4.7 Results on Hanabi
      • 7.5 Related Work
        • 7.5.1 Learning to Communicate
        • 7.5.2 Research on Hanabi
        • 7.5.3 Belief State Methods
      • 7.6 Conclusion & Future Work
  • III Learning to Reciprocate
    • 8. Learning with Opponent-Learning Awareness
      • 8.1 Introduction
      • 8.2 Related Work
      • 8.3 Methods
        • 8.3.1 Naive Learner
        • 8.3.2 Learning with Opponent Learning Awareness
        • 8.3.3. Learning via Policy gradient
        • 8.3.4 LOLA with Opponent modeling
        • 8.3.5 Higher-Order LOLA
      • 8.4 Experimental Setup
        • 8.4.1 Iterated Games
        • 8.4.2 Coin Game
        • 8.4.3 Training Details
      • 8.5 Results
        • 8.5.1 Iterated Games
        • 8.5.2 Coin Game
        • 8.5.3 Exploitability of LOLA
      • 8.6 Conclusion & Future Work
    • 9. DiCE: The Infinitely Differentiable Monte Carlo Estimator
      • 9.1 Introduction
      • 9.2 Background
        • 9.2.1 Stochastic Computation Graphs
        • 9.2.2 Surrogate Losses
      • 9.3 Higher Order Gradients
        • 9.3.1 Higher Order Gradient Estimators
        • 9.3.2 Higher Order Surrogate Losses
        • 9.3.3. Simple Failing Example
      • 9.4 Correct Gradient Estimators with DiCE
        • 9.4.1 Implement of DiCE
        • 9.4.2 Casuality
        • 9.4.3 First Order Variance Reduction
        • 9.4.4 Hessian-Vector Product
      • 9.5 Case Studies
        • 9.5.1 Empirical Verification
        • 9.5.2 DiCE For multi-agent RL
      • 9.6 Related Work
      • 9.7 Conclusion & Future Work
  • Reference
    • Reference
  • After
    • ๋ณด์ถฉ
    • ์—ญ์ž ํ›„๊ธฐ
Powered by GitBook
On this page

Was this helpful?

  1. II Learning to Communicate
  2. 7. Bayesian Action Decoder
  3. 7.4 Experiments and Results

7.4.4 Beliefs in Hanabi

Hanabi์˜ belief๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ฐ€์žฅ ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ftpubf^{\mathrm{pub}}_{t}ftpubโ€‹๋Š” ๋ฑ์— ๋‚จ์•„์žˆ๋Š” ์นด๋“œ์— ๋Œ€ํ•œ ์ •๋ณด candidates vector C CC์™€, hint mask HM HMHM์€ nNhร—(NcolorNrank+1) nN_h \times (N_{\mathrm{color}}N_{\mathrm{rank}}+1)nNhโ€‹ร—(Ncolorโ€‹Nrankโ€‹+1) ํฌ๊ธฐ์˜ binary matrix๋กœ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.(ํŒจ๊ฐ€ ์—†๋‹ค๋Š” ์ •๋ณด +1๊นŒ์ง€)์ด๋Š” ์ฃผ์–ด์ง„ ์Šฌ๋กฏ์—์„œ ํ”Œ๋ ˆ์ด์–ด๊ฐ€ ์ง€๊ธˆ๊นŒ์ง€ ํžŒํŠธ์— ๋”ฐ๋ผ ํŠน์ • ์นด๋“œ๋ฅผ ๋“ค๊ณ  ์žˆ์„ ์ˆ˜ ์žˆ๋Š” ๊ฒฝ์šฐ 1, ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด 0์ธ matrix์ž…๋‹ˆ๋‹ค. ์Šฌ๋กฏ์€ public state์— ๋Œ€ํ•œ private state space f[i]f[i]f[i]๋กœ ๋ถ€ํ„ฐ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. basic belief B0B^0B0๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค.

B0(f[i])=P(f[i]โˆฃfpub)โˆC(f)ร—HM(f[i]) B^0(f[i]) = P(f[i]|f^{\mathrm{pub}})\propto C(f) \times HM(f[i])B0(f[i])=P(f[i]โˆฃfpub)โˆC(f)ร—HM(f[i])

์ด๋ฅผ "V0 belief" ๋ผ๊ณ  ๋ถ€๋ฅผ๊ฑด๋ฐ, ์ด๋Š” ๊ทธ ์นด๋“œ์— ๋Œ€ํ•œ public available information์œผ๋กœ๋งŒ ์ด๋ฃจ์–ด์ ธ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฒˆ ์‹คํ—˜์—์„œ baseline agent๋Š” basic belief ๋ฅผ ๋ฐ›๋Š” ๊ฒƒ์„ ์ค‘์ ์œผ๋กœ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด ์ •๋ณด๋Š” ์‹ค์ œ ์‚ฌ๋žŒ์ด ์ด์šฉํ•˜๊ธฐ์—๋Š” ๋ชจ๋“  ํžŒํŠธ๋ฅผ ๊ธฐ์–ตํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™๊ธฐ ๋•Œ๋ฌธ์— ๋ฌธ์ œ๊ฐ€ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  basic belief๋Š” ๋‹ค๋ฅธ ์ •๋ณด์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ํ†ตํ•œ ์–ด๋–ค ๊ฒฐ๊ณผ๋ฌผ์„ ๋‚ด๋†“์€ ๊ฒƒ์— ๋Œ€ํ•œ ๊ณ ๋ ค๋ฅผ ํ•˜์ง€ ์•Š์€ ์ฑ„๋กœ ์ด๋ฃจ์–ด์ง‘๋‹ˆ๋‹ค. ์ด๋ฅผ BAD์—์„œ๋Š” self-consistent belief์™€ (7.3.11) ์ˆ˜์‹์—์„œ ์ด๋ฃจ์–ด์กŒ๋˜ noisy sampling์„ ํ†ตํ•ด ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค.

๊ฐ„๋‹จํ•œ ๊ฐ ์นด๋“œ์— ๋Œ€ํ•œ belief๋Š” ๋‹ค์Œ์ฒ˜๋Ÿผ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

B0(f[i])โˆC(f)ร—HM(f[i])ร—L(f[i]) B^0(f[i]) \propto C(f) \times HM(f[i]) \times \mathcal{L}(f[i])B0(f[i])โˆC(f)ร—HM(f[i])ร—L(f[i])

B0(f[i])=C(f)ร—HM(f[i])ร—L(f[i])โˆ‘gC(g)ร—HM(f[i])ร—L(f[i]) B^0(f[i]) = \frac{C(f) \times HM(f[i]) \times \mathcal{L}(f[i])}{\sum_gC(g) \times HM(f[i]) \times \mathcal{L}(f[i])} B0(f[i])=โˆ‘gโ€‹C(g)ร—HM(f[i])ร—L(f[i])C(f)ร—HM(f[i])ร—L(f[i])โ€‹

=ฮฒiC(f)ร—HM(f[i])ร—L(f[i])=\beta_i {C(f) \times HM(f[i]) \times \mathcal{L}(f[i])} =ฮฒiโ€‹C(f)ร—HM(f[i])ร—L(f[i])

์•„๋ž˜ ๋‘ term์€ probability์ฒ˜๋Ÿผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ normalization์„ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค.

๋‹ค์Œ์œผ๋กœ๋Š” iterative belief update๋ฅผ ์ง„ํ–‰ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.

Bk+1(f[i])=โˆ‘f[โˆ’i]Bk(f[โˆ’i])P(f[i]โˆฃf[โˆ’i],fโ‰คtpub,uโ‰คta,ฯ€^โ‰คt)\mathcal{B}^{k+1}(f[i]) = \sum_{f[-i]}\mathcal{B}^k(f[-i])P(f[i]|f[-i],f^\mathrm{pub}_{\leq t},u^a_{\leq t}, \hat{\pi}_{\leq t})Bk+1(f[i])=โˆ‘f[โˆ’i]โ€‹Bk(f[โˆ’i])P(f[i]โˆฃf[โˆ’i],fโ‰คtpubโ€‹,uโ‰คtaโ€‹,ฯ€^โ‰คtโ€‹)

=โˆ‘g[โˆ’i]Bk(g[โˆ’i])ฮฒi(C(f)โˆ’โˆ‘jโ‰ i1(g[j]=f))M(f[i]) =\sum_{g[-i]}\mathcal{B}^k(g[-i])\beta_i(C(f)- \sum_{j \neq i} \bm{1}(g[j] = f)) M(f[i]) =โˆ‘g[โˆ’i]โ€‹Bk(g[โˆ’i])ฮฒiโ€‹(C(f)โˆ’โˆ‘j๎€ =iโ€‹1(g[j]=f))M(f[i])

์ด ๋•Œ M(f[i])=HM(f[i])ร—L(f[i]) M(f[i]) = HM(f[i]) \times \mathcal{L}(f[i])M(f[i])=HM(f[i])ร—L(f[i])์ž…๋‹ˆ๋‹ค. ์ด๋•Œ, Bk\mathcal{B}^kBk์— ๋Œ€ํ•ด factorized ํ•˜๊ฒŒ ๊ทผ์‚ฌํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

Bk+1(f[i])=โˆ‘g[โˆ’i]Bk(g[โˆ’i])ฮฒi(C(f))โˆ’โˆ‘jโ‰ i1(g[j]=f))M(f[i]) B^{k+1}(f[i]) = \sum_{g[-i]}\mathcal{B}^k(g[-i])\beta_i(C(f))- \sum_{j \neq i} \bm{1}(g[j] = f)) M(f[i])Bk+1(f[i])=โˆ‘g[โˆ’i]โ€‹Bk(g[โˆ’i])ฮฒiโ€‹(C(f))โˆ’โˆ‘j๎€ =iโ€‹1(g[j]=f))M(f[i])

=โˆ‘g[โˆ’i]โˆjโ‰ iBk(g[โˆ’i])ฮฒi(C(f))โˆ’โˆ‘jโ‰ i1(g[j]=f))M(f[i])= \sum_{g[-i]}\prod_{j \neq i}\mathcal{B}^k(g[-i])\beta_i(C(f))- \sum_{j \neq i} \bm{1}(g[j] = f)) M(f[i])=โˆ‘g[โˆ’i]โ€‹โˆj๎€ =iโ€‹Bk(g[โˆ’i])ฮฒiโ€‹(C(f))โˆ’โˆ‘j๎€ =iโ€‹1(g[j]=f))M(f[i])

โ‰ƒฮฒiโˆ‘g[โˆ’i]โˆjโ‰ iBk(g[โˆ’i])ฮฒi(C(f))โˆ’โˆ‘jโ‰ i1(g[j]=f))M(f[i])\simeq \beta_i\sum_{g[-i]}\prod_{j \neq i}\mathcal{B}^k(g[-i])\beta_i(C(f))- \sum_{j \neq i} \bm{1}(g[j] = f)) M(f[i])โ‰ƒฮฒiโ€‹โˆ‘g[โˆ’i]โ€‹โˆj๎€ =iโ€‹Bk(g[โˆ’i])ฮฒiโ€‹(C(f))โˆ’โˆ‘j๎€ =iโ€‹1(g[j]=f))M(f[i])

๋งˆ์ง€๋ง‰ term์€ sample์„ ํ‰๊ท ํ•œ ํ›„ normalizing์„ ํ•ด ๊ฒฐ๊ณผ๊ฐ’๊ณผ๋Š” ์ฐจ์ด๊ฐ€ ์ƒ๊ธฐ์ง€๋งŒ, ์ด๋Š” sample๋งˆ๋‹ค normalizing์„ํ•˜๊ณ , ํ‰๊ท ์„ ๋‚ด๋Š” ๊ฒƒ๋ณด๋‹ค ์ข€ ๋” ๋‹ค๋ฃจ๊ธฐ ์‰ฌ์šด์ ์„ ์ด์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

์ด๋Š” product-sum์„ ํ†ตํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Bk+1(f[i])โ‰ƒฮฒi(C(f)โˆ’โˆ‘g[โˆ’i]โˆjโ‰ iBk(g[j])โˆ‘jโ‰ i1(g[j]=f))M(f[i]) \mathcal{B}^{k+1}(f[i]) \simeq \beta_i(C(f) - \sum_{g[-i]}\prod_{j\neq i} \mathcal{B}^k(g[j]) \sum_{j \neq i } \bm{1}(g[j] = f)) M(f[i])Bk+1(f[i])โ‰ƒฮฒiโ€‹(C(f)โˆ’โˆ‘g[โˆ’i]โ€‹โˆj๎€ =iโ€‹Bk(g[j])โˆ‘j๎€ =iโ€‹1(g[j]=f))M(f[i])

=ฮฒi(C(f)โˆ’โˆ‘jโ‰ iโˆ‘gBk(g[j])1(g[j]=f))M(f[i]) = \beta_i(C(f) - \sum_{{j\neq i} }\sum_g\mathcal{B}^k(g[j]) \bm{1}(g[j] = f)) M(f[i])=ฮฒiโ€‹(C(f)โˆ’โˆ‘j๎€ =iโ€‹โˆ‘gโ€‹Bk(g[j])1(g[j]=f))M(f[i])

=ฮฒi(C(f)โˆ’โˆ‘jโ‰ iBk(g[j]))M(f[i]) = \beta_i(C(f) - \sum_{{j\neq i} }\mathcal{B}^k(g[j]) ) M(f[i])=ฮฒiโ€‹(C(f)โˆ’โˆ‘j๎€ =iโ€‹Bk(g[j]))M(f[i])

โˆ(C(f)โˆ’โˆ‘jโ‰ iBk(g[j]))M(f[i])\propto(C(f) - \sum_{{j\neq i} }\mathcal{B}^k(g[j]) ) M(f[i])โˆ(C(f)โˆ’โˆ‘j๎€ =iโ€‹Bk(g[j]))M(f[i])

๊ทธ๋Ÿฌ๋ฏ€๋กœ self consistent belief๋ฅผ sampling ์—†์ด ๋‹ค์Œ์ฒ˜๋Ÿผ ๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Bk+1(f[i])โˆ(C(f)โˆ’โˆ‘jโ‰ iBk(g[j]))ร—HM(f[i]) \mathcal{B}^{k+1}(f[i]) \propto(C(f) - \sum_{{j\neq i} }\mathcal{B}^k(g[j]) ) \times HM(f[i])Bk+1(f[i])โˆ(C(f)โˆ’โˆ‘j๎€ =iโ€‹Bk(g[j]))ร—HM(f[i])

๊ทธ๋ฆฌ๊ณ  belief๊ฐ€ ์ˆ˜๋ ดํ•˜๊ฑฐ๋‚˜ iteration์„ ์ตœ๋Œ€๋กœ ํ–ˆ์„ ๋•Œ, ์ด๋ฅผ V1 belief๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค. ์ด๋Š” Bayesian probability์— ์˜ํ•˜์—ฌ ๋งŒ๋“ค์–ด์กŒ์ง„ ์•Š์ง€๋งŒ, ๋‹ค๋ฅธ์นด๋“œ์™€ ํžŒํŠธ์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ํ†ตํ•ด ๋งŒ๋“ค์–ด์กŒ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ iteration์„ ํ• ๋•Œ๋งˆ๋‹ค HM์˜ slot์€ ๋‹ค๋ฅธ slot๋“ค์— ์˜ํ•ด ๋“ค์–ด๊ฐˆ candidate๊ฐ€ ์ค„์–ด๋“œ๋Š” ์ถ”๋ก ์ด ์ด๋ค„์ง‘๋‹ˆ๋‹ค.

๊ฐ™์€ algorithm์ด์ง€๋งŒ L\mathcal{L}L์„ ํฌํ•จํ•˜๋Š” algorithm์„ BAD๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

BB0(f[i])โˆC(f)ร—HM(f[i])ร—L(f[i]) BB^0(f[i]) \propto C(f) \times HM(f[i]) \times \mathcal{L}(f[i]) BB0(f[i])โˆC(f)ร—HM(f[i])ร—L(f[i])

BB1(f[i])โˆ(C(f)โˆ’โˆ‘jโ‰ iBk(f[j]))ร—HM(f[i])ร—L(f[i]) BB^1(f[i]) \propto (C(f)-\sum_{j \neq i }B^k(f[j])) \times HM(f[i]) \times \mathcal{L}(f[i]) BB1(f[i])โˆ(C(f)โˆ’โˆ‘j๎€ =iโ€‹Bk(f[j]))ร—HM(f[i])ร—L(f[i])

์‹ค์ œ๋กœ V2 belief๋Š” ์•ˆ์ •์„ฑ์„ ์œ„ํ•ด Bayesian Belief์™€ V1 belief์˜ ๋ณด๊ฐ„๋ฒ•์„ ํ†ตํ•ด ๊ตฌํ•ฉ๋‹ˆ๋‹ค.

V2=(1โˆ’ฮฑ)BB+ฮฑV1ย ย ,ฮฑ=0.01ย ย orย ย 0.1 V2 = (1-\alpha) BB+ \alpha V1 \ \ , \alpha = 0.01 \ \ \mathrm{or} \ \ 0.1V2=(1โˆ’ฮฑ)BB+ฮฑV1ย ย ,ฮฑ=0.01ย ย orย ย 0.1

Previous7.4.3 Observations and ActionsNext7.4.5 Architecture Details for Baselines and Method

Last updated 4 years ago

Was this helpful?