buyer agent decision process based on automatic fuzzy rules generation methods roi arapoglou, kostas...
TRANSCRIPT
Buyer Agent Decision Process Based on Automatic Fuzzy Rules Generation
Methods
Roi Arapoglou, Kostas Kolomvatsos, Stathes Hadjiefthymiades
Pervasive Computing Research Group, Department of
Informatics and Telecommunications
University of Athens, Greece
WCCI – FUZZ 2010
Barcelona - Spain
Outline
2
Introduction Market Members – Scenario Buyer Behavior – Decision Process Buyer Fuzzy Logic System Fuzzy Rules Generation Results
Introduction
3
Intelligent Agents Autonomous software components Represent users Learn from their owners
Electronic Markets Places where entities not known in advance can
negotiate for the exchange of products Fuzzy Logic
Algebra based on fuzzy sets Deals with incomplete or uncertain information Enhance the knowledge base of agents
Market Members - Scenario
4
Buyers Sellers Middle entities (matchmakers, brokers, market
entities) Intelligent Agents may represent each of these entities
Scenario Modeled as a finite-horizon Bargaining Game No knowledge about the characteristics of the opponent
(i.e., the other side) is available
Buyer Behavior – Decision process (1/2)
5
The buyer stays in the game for a specific number of rounds
Profit A Utility Function is used , where V is the buyer valuation and p is
the product price The smaller the price is the greater the profit
becomes Pricing Function , where p0 is an initial price, V is the
valuation, x is the number of the proposal, Tb is the deadline and k is a policy factor (k>1:patient, k<1:aggressive, k=1:neutral)
pVU b
k1b0
bt )T(xVpp
Buyer Behavior – Decision process (2/2) Receives proposals and accepts or rejects them
making its own proposals Utilizes a reasoning mechanism based on FL The mechanism results the value of the
Acceptance Degree (AD) The reasoning mechanism is based on the
following parameters: Relevance factor (r) Price difference (d) Belief about the expiration of the game (b) Time difference (t) Valuation (V)
6
Buyer Fuzzy Logic System (1/2) Architecture
Contains a set of Fuzzy rules Rules are automatically generated based on
experts dataset
7
Buyer Fuzzy Logic System (2/2) Advantages of the automatic Fuzzy rules
generation Mainly, it does not require a lot of time in the
developer side It does not require experience in FL rules definition It uses simple numbers representing values of
basic parameters Fuzzy rules are automatically tuned
8
Fuzzy Rules Generation (1/2) Clustering techniques are used Algorithms:
K-means Fuzzy C-means (FCM) Subtractive clustering Nearest Neighborhood Clustering (NNC)
Every cluster corresponds to a Fuzzy rule Example If is a cluster center the rule is:
9
)x,...,x,(x*x *n
*2
*1
*nn
*1-n1-n
*22
*11 x is xTHEN x is x AND... x is x ANDx is x IF
Fuzzy Rules Generation (2/2) Additional techniques
Learning from Examples (LFE) Modified Learning from Examples (MLFE)
Templates for membership functions are defined
Dataset They describe the policy that the buyer should
have, concernig the acceptance of a proposal 108 rows of data Each row contains data for r, d, b, t, and V
10
Results (1/3) Fuzzy rule base creation time
Usage of the generated Fuzzy rule base in a BG We use the following parameters
We examine the Joint Utility in seven agreement zones (theoretic maximum equal to 0.25)
, (1) where P* is the agreement price, C is the seller cost and V is the buyer valuation
11
Algorithm Rule Base creation time (ms)Subtractive 35FCM 2560K-Means 25LFE 20MLFE 25NNC 20
Buyer Parameters Seller Parameters
Initial Price 100 MUs[1] Cost 250 MUs
Valuation 255 MUs Initial Profit 250 MUs[1] MU = Monetary Unit
2
**
C)(V
)P(VC)(PJU
(1)D. Zeng & K. Sycara, ‘Bayesian Learning in Negotiation’, International Journal of Human-Computer Studies, vol(48), no 1, 1998, pp. 125-141.
Results (2/3) Agreement zones
Numerical results
12
Buyer Valuation Agreement Zone255 MUs 5 MUs260 MUs 10 MUs270 MUs 20 MUs300 MUs 50 MUs500 MUs 250 MUs700 MUs 450 MUs
1000 MUs 750 MUs
Scenario No
Agreement Zone
Average JU Maximum JU Algorithm
1 5 MUs 0.08 0.24 FCM, K-Means2 10 MUs 0.14 0.24 FCM, K-Means3 20 MUs 0.16 0.21 LFE4 50 MUs 0.24 0.247 FCM, K-Means5 250 MUs 0.238 0.24 MLFE6 450 MUs 0.208 0.21 MLFE7 750 MUs 0.17 0.172 MLFE
Results (3/3) Performance of algorithms in the BG
13
Algorithm Agreements Percentage Average JUSubtractive 92% 0.217
FCM 69% 0.219K-Means 69% 0.202
LFE 57% 0.223MLFE 85% 0.244NNC 86% 0.244
Algorithm Average AD ValueSubtractive 80.96
FCM 68.91K-Means 62.84
LFE 72.65MLFE 74.52NNC 76.58