Predicting Agents Tactics in Automated Negotiation

1
Predicting Agents Tactics in Automated Negotiation

• Chongming Hou
• Knowledge Media Institute
• The Open University, UK

2
Outline

• Negotiation
• Decision Functions
• Learning Approach
• Negotiation Heuristics
• Performance
• Discussion
• Conclusion

3
Negotiation

• Jointly exploring the space of possible deals to
  reach an agreement
• Negotiation
• Language and ontologies
• Agent communication language (ACL),KQML
• Content language SL,RDF
• Protocol
• Ask, broker, bid, respond, agree
• Decision
• Policies, strategies, tactics

4
Negotiation Decision Model

• Two party adopting opposing roles, such as buyer
  and seller
• Alternating offers

5
Related Work

• Game theory
• Assumes common knowledge and perfect rationality
• Decision functions
• Time-dependent tactics
• Resource-dependent tactics
• Behaviour-dependent tactics

6
Decision Parameters

• Reservation values for seller and buyer mins,
  maxb
• Initial offer level k
• Deadline (time available) tmax
• Rate of concession ß/µ/?/d,
• Number of negotiating agents N(t)
• Communication cost Xs?b

7
Decision Functions

• Time
• Polynomial
• Exponential
• Resource ,
• Behaviour
• AvgTFT RelTFT
• RndTFT

8
Time Dependent Functions
9
Motivation

• Related work on decision functions
• Require prior probability distribution of
  opponents deadline and reservation value
• Require knowledge of opponents tactic type
• Research questions
• How can an agent learn the opponents tactics?
• How can knowledge of the opponents tactic be
  used to achieve the optimal outcome for the
  predicting agent?

10
Learning Approach

• Use nonlinear regression to learn an opponents
  tactic
• Discriminate which of the three families of
  tactics the opponent has adopted
• Predict the opponents future offers
• Identify the opponents tactic type
• Estimate the opponents reservation value and
  deadline

11
Negotiation Heuristics

• Three sets of heuristics
• Boulware, Conceder and Resource tactics
• Example Opponent is a Time-Dependent (min SSR),
  Conceder (ßgt1) tactic
• Accept the opponents offer if
• offer(tn) - max lt 0.5 ln(max), or
• a gt 0.90 to 0.98, or
• tn - tmax lt 1, or
• offer(tn) - offer(tn-1) lt max 1

12
Performance

• Time-dependent
• Conceder ß 20, 30, 40, Linear ß 1
• Boulware ß 0.025, 0.1, 0.2
• Resource-dependent
• Impatient µ 1, Steady µ 2, 3, 5,
  Patient µ 6, 8, 10
• Behaviour-dependent
• Average Tit-For-Tat ? 1, 2, 3, 5, 6, 8, 10,
• Relative TFT d 1, 2, 3, 5, 6, 8, 10,
• Random TFT d 1, 2, 3, 5, 6, 8, 10, M 1, 3

13
Performance

• Performance function
• Pairwise comparison of 90 paired conditions
• Buyer Conceder, Linear, Boulware, Impatient,
  Steady, Patient, Average TFT, Relative TFT, and
  Random TFT
• Seller Conceder, Linear, Boulware, Impatient,
  Steady, Patient, Average TFT, Relative TFT,
  Random TFT, and Prediction

14
Performance Time

• The performance of the sellers learning
  mechanism against the time-dependent buyer

15
Performance Resource

• The performance of the sellers learning
  mechanism against the resource-dependent buyer

16
Performance Behaviour

• The performance of the sellers learning
  mechanism against the Tit-For-Tat dependent buyer

17
Discussion

• Withdraw from pointless negotiation
• Withdraw from unprofitable negotiation
• Avoid breakdown and make best deal

18
Conclusion

• Predict opponents behaviour
• Effectively
• Detect the pointless negotiation
• Balance future gain and cost
• Avoid breakdown of negotiation whilst making a
  the best deal at the opponents reservation value
• Online
• Without prior knowledge about opponent
• Based only on the opponents previous offers
• These characteristics give this method
  significant practical value in negotiation