DAgents Resource Allocation Model

1
DAgents Resource Allocation Model

• A Market-Based Model for Resource Allocation in
  Agent Systems

2
Introduction to DAgents

• Mobile Agent system developed at Dartmouth
  College
• Sponsored by DARPA (Defense Advanced Research
  Projects Agency) and several other military
  branches.
• Used mobile agents on mobile devices for a test
  scenario.

3
Problems with Agent Systems

• Why should a host allow a mobile agent to
  execute?
• Hosting a mobile agent exposes host to risk of
  malicious agents
• No incentive to host agents.
• They cannot be trusted.
• They are extra processes the host must
  accommodate.

4
How do we provide incentive?

• We need something to provide an incentive

• Money! Money! Money! MONEY!
• Money talks, just ask Donald Trump

5
Market Model

• We are paid for the labour and/or skills we
  provide our employers or clients.
• Payments money spending potential
• Why not apply same theory to agents?
• Hosts provide a service to agents.
• Agents provide services to each other and users.

6
Market System for DAgents

• Agents purchase resource usage from hosts, or
  other agents.
• Hosts then redistribute revenue to users.
• Users spend on launching own agents.

7
Currency

• Agent currency can range from symbolic

• to real world tender.

8
Secure Transactions

• Like real world trade, require verifiable
  currency.
• DAgents implements a mediation protocol.
  (Similar to SafeTrader on Trademe).
• Verification of the currency can verify the
  spender.

9
Market as a Security Measure

• Protecting agents from agents tractable.
• Protecting hosts from agents tractable.
• Protecting agents from hosts difficult.
• Tampering with an agent may dissuade agent from
  returning, thus discouraging income. Agents
  will not visit and host will not be able to
  launch agents.

10
Arbiter

• Arbiter provides a trusted third party (i.e. CA
  for HTTPS)
• Arbiter collects collateral and information from
  agents.
• If both agents are satisfied, collateral is
  returned.
• Otherwise, Arbiter retains collateral until
  resolved.

11
Resource Managers

• Resource managers handle resource allocation on
  behalf of the host.
• Allow system administrators to fine tune access
  to system resources.

12
Allocation Mechanism

• Purpose of the mechanism is to prioritize rather
  than generate revenue.
• Hosts solicit bids from agents to assess demand
  on system.
• Assumes agent uses currency to complete its
  tasks.
• Bid strategy transformed over domain common to
  all bidding agents.

13
Allocation Mechanism

• Host calculates sum of competing bids and
  allocates a proportion of resources to each
  agent.
• Provable that there is one bidding level
  satisfying all agents.
• Game has Nash Equilibrium.

14
Experimental Results

• Run on Swarm
• Performance gradually decays as latency
  increases, despite having older information.
• If load is too high, richer agents are given
  priority.

• Karl Marx would not approve!

15
Experimental Results

• Price is log-normally distributed, so price
  will be mostly stable with brief price spikes.
• Agents will continue to enter system, so price
  will be high after spike.
• More desirable to know how an agent will perform,
  so further work includes call options and agent
  utility Bredin, Kotz, Rus1999

16
Further Research and Papers

• Utility Driven Mobile Agent Scheduling
• Jonathan Bredin, David Kotz and Daniela Rus
  1999
• http//agent.cs.dartmouth.edu/papers/bredindeman
  d.pdf
• A Game-Theoretic Formulation of Multi-Agent
  Resource Allocation
• Jonathan Bredin, Rajiv T. Maheswaran, Cagri
  Imer, Tamer Basar, David Kotz and Daniela Rus
  2000
• http//agent.cs.dartmouth.edu/papers/bredingame.
  pdf

17
Appendix

• Note I didnt actually ask, but I would assume
  so.