Mechanisms to Promote Cooperation in Decentralized Services
1. Outline System Model Cooperation Mechanisms Results Conclusions
Mechanisms to promote cooperation
in decentralized services
E. del Val M. Rebollo V. Botti
Univ. Politecnica de Valencia (Spain)
EUMAS ’12
Dublin, December 2012
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
2. Outline System Model Cooperation Mechanisms Results Conclusions
Promoting Cooperation
Motivation
There are scenarios in decentralized systems in which cooperation
plays a central role
agents connected in networks
bounded rationality
heterogeneous, self-interested agents
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
3. Outline System Model Cooperation Mechanisms Results Conclusions
Our Proposal
The challenge
Obtain an emergent, cooperative global behavior even when
cooperators are a minority, from local decisions.
What is done. . .
a network structure that ensures navigation and efficiency
structural changes to isolate undesired agents
incentives to promote cooperation
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
4. Outline System Model Cooperation Mechanisms Results Conclusions
Outline
1 Outline
2 System Model
3 Cooperation Mechanisms
4 Results
5 Conclusions
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
5. Outline System Model Cooperation Mechanisms Results Conclusions
System Model
Definition (Open Service-Oriented MAS)
(A, L), where A = {ai , ..., an } is a finite set of autonomous agents
that are part of the system, and L ⊆ A × A is the set of links,
where each link (ai , aj ) ∈ L
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
6. Outline System Model Cooperation Mechanisms Results Conclusions
System Model
Definition (Agent)
is a tuple (Si , Ni , sti ) where:
Si = {s1 , . . . , sl } is the set of semantic service descriptions of
the services provided by the agent (WSDL);
Ni is the set of neighbors of the agent,
Ni ⊆ A − {ai } : ∀aj ∈ Ni , ∃(ai , aj ) ∈ L, and |Ni | > 0. It is
assumed that |Ni | |A|;
sti is the internal state of the agent.
πi : sti → Ni is the neighbor selection function that
determines the most promising neighbor to provide a service;
ρi : sti → Ψ is the adaptation selection function where Ψ is
the set of finite adaptation actions of the agent.
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
7. Outline System Model Cooperation Mechanisms Results Conclusions
Network Creation
probabilistic relations baed on homophily (assortativity,
similarity)
two agents are similar if they provide similar services
H = αHv + (1 − α)Hs
growing network structure (exponential degree distribution)
certain characteristics of a small-world network (short paths,
clustering)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
8. Outline System Model Cooperation Mechanisms Results Conclusions
Service Discovery
navigation problem in networks (it relies in the agent
cooperation)
hill climbing (greedy) method
|Nj |
H(aj , at )
π(at ) = argmax 1 − 1 −
H(an , at )
aj ∈Ni
an ∈Ni
agents pass the query until the desire service is found
it is a problem with self–interested agents
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
9. Outline System Model Cooperation Mechanisms Results Conclusions
Social Plasticity
capacity to change
relations as times passes
link utility decays with
time
depends on the queries aj
forwards
1
D(aj ) = 1+e −γ
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
10. Outline System Model Cooperation Mechanisms Results Conclusions
Social Plasticity
when an agent breaks a link, a substitute must be found
(maintain the network structure)
criteria
neighbor of neighbor
a similar agent to the previous one
rewire links has not a cost
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
11. Outline System Model Cooperation Mechanisms Results Conclusions
Social Plasticity
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
12. Outline System Model Cooperation Mechanisms Results Conclusions
Social Plasticity
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
13. Outline System Model Cooperation Mechanisms Results Conclusions
Incentives
each action implies a cost (ask for a service and forward)
a reward is obtained if the service is found
rewards are provided by the system
agents imitated the strategy of successful neighbors
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
14. Outline System Model Cooperation Mechanisms Results Conclusions
Incentives
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
15. Outline System Model Cooperation Mechanisms Results Conclusions
Combining social plasticity and incentives
Both strategies promote cooperation in general
but it is not enough if non–cooperative agents has a high
degree
network broken in isolated parts
rewire cost –> not affordable for some agents
payoff not enough to promote cooperation
the combined model
1 incentives to change the behavior of non–cooperatives
2 rewire links if it fails
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
16. Outline System Model Cooperation Mechanisms Results Conclusions
Experiment Design
Configuration
1,000 agents, 10 different networks
100 steps to forward a query, snapshots after 5,000 queries
varying the initial prop. of collaborators
Strategies
Social plasticity (SP)
Incentives
Reinforcement Learning (RL) using WOLF
Game-theory approach, using Prisoner’s Dilemma (PD)
Incentives + Social Plasticity (I+SP)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
17. Outline System Model Cooperation Mechanisms Results Conclusions
Measures
proportion of collaborator / non–collaborators
average path length (better if smaller)
search failures due to non–collaboration
search success (including TTL)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
18. Outline System Model Cooperation Mechanisms Results Conclusions
60% of collaborators (num and path length)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
19. Outline System Model Cooperation Mechanisms Results Conclusions
40% of collaborators (num and path length)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
20. Outline System Model Cooperation Mechanisms Results Conclusions
60% of collaborators (failures and success)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
21. Outline System Model Cooperation Mechanisms Results Conclusions
40% of collaborators (failures and success)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services
22. Outline System Model Cooperation Mechanisms Results Conclusions
Conclusions
improving cooperation to solve navigation problem in networks
applied to decentralized service management
combination of structural changes and incentives
improves the performance when non-collaborators are
’important’ in the network
works by imitation: a core of collaborators is needed
a guess: the size of the core depends on network
characteristics (percolation, efficience, centrality coefficient)
M. Rebollo et al. (UPV) EUMAS’12
Mechanisms to promote cooperation in decentralized services