A N EW P ERSPECTIVE OF T RUST THROUGH M ULTI - A TTRIBUTE A UCTIONS - PowerPoint PPT Presentation

A N EW P ERSPECTIVE OF T RUST THROUGH M ULTI - A TTRIBUTE A UCTIONS Ferran Torrent-Fontbona Albert Pla Beatriz López

Auctions in workflow management systems  Auctions allow an optimal allocation for just-in-time  Competitive market  Special domains: – Production under demand / Supply chain under demand – Handling unexpected tasks (provoked by faults) – Unknown resource status – Outsourced tasks Workflow Agent A Resource Type A AUCTION! Resource Resource Resource ··· Agent n Agent 1 Agent 2 November 13, 2015 2/19

Multi-dimensional allocation problem  Production process managers are not only concerned by costs  Workflow managers are concerned about multiple attributes: – Economic costs – Product quality – Delivery times – Environmental footprint – Licenses / ISO standardizations – … November 13, 2015 3/19

Multi-dimensional allocation problem Multi-criteria allocation Multi-attribute auctions problem November 13, 2015 4/19

Trust motivation  Misdelivered tasks involve: November 13, 2015 5/19

Trust motivation  Misdelivered tasks are due: – Cheating behaviors – Involuntary errors • Bidders may not be able to accurately estimate their abilities November 13, 2015 6/19

Trust motivation  Cheating agents: – Incentive Compatible Mechanism • Vickrey Based Auction (VCG Payment rule) • …  Involuntary errors and misestimating the abilities – Trust & Reputation based auctions • Porter’s auction ( uni-attribute) • Ramchurn’s auction (uni-attribute) • …  No solution integrating Incentive compatibility, trust & multi-attribute November 13, 2015 7/19

Multi-attribute resource/task allocation Auctions Uni- Multi- Attribute Attribute Vickrey Position Score Flexible Uncertain Auctions Auctions Auctions Attribute Delivery Auctions Vickrey Che’s Parke’s English 1 attribute + 1 N attributes + GSP VCG Auction PERA n attributes VCG trust attribute Auction Auctions M trust attributes Single Unit First-score Trust Extended- Google PPC De Smet PUMAA Vickrey Auction VCG ∅ Discriminatory Second-Score Porter’s Fault Mahr fair-PUMAA MU Vickrey Auction Tolerant Auction Non- Second-Preferred Trust-Based … Discriminatory Offer Auctions MU Vickrey Suitable for the considered task allocation Use of trust or reputation Not suitable for the considered task allocation Multi-attribute (excluding trust) November 13, 2015 8/19

Methodology 1. Call for proposals (CFP) 2. Bidding 3. Winner determination problem (WDP) 4. Payment 5. Trust learning November 13, 2015 9/19

1. Call for proposals  An auctioneer 𝐵 0 needs to allocate a task 𝑈 0 with a set of attributes 𝑏 1 , … , 𝑏 𝑜  It Sends a call for proposals (CFP) to all the bidders – Specifies the task – Specifies the attribute to evaluate – Specifies the evaluation function Bidder Bidder 𝐷𝐺𝑄 1 2 0 , 𝑊 ∙ 0 , … , 𝑏 𝑜 Auctioneer = 𝑈 0 , 𝑏 1 Bidder i November 13, 2015 10/19

2. Bidding  Bidders evaluate the CFP and submit the bids with the corresponding attributes 𝐶 𝑗 = 𝑐 𝑗 , 𝑢 𝑗 , 𝑓 𝑗  Each bidder submits the bid that is expected to maximize its utility Bidder Bidder Auctioneer 1 2 Bidder i November 13, 2015 11/19

3. Winner determination problem  Inclusion of trust in the valuation of the bid – One trust attribute per checkable attribute = 𝜑 𝑈 0 − 𝑊 𝑐 𝑗 , 𝑢 𝑗 𝑢 , 𝑓 𝑗 𝑢 , 𝜐 𝑗,𝑠 𝑓 𝑣 𝑈 0 , 𝑐 𝑗 , 𝑢 𝑗 , 𝑓 𝑗 , 𝜐 𝑗,𝑠 𝑓 𝜐 𝑗,𝑠 𝜐 𝑗,𝑠  WDP consists of finding the bid that minimizes the evaluation function 𝑊 𝑐 𝑗 , 𝑢 𝑗 𝑢 , 𝑓 𝑗 min 𝑓 𝜐 𝑗,𝑠 𝜐 𝑗,𝑠 𝑗 Bidder Bidder Auctioneer 1 2 Evaluate (Optimize € , kWh, time, trust…) Bidder i Bid 1 Bid 4 Bid 2 … Bid 3 Bid n November 13, 2015 12/19

4. Payment  Conditional Vickrey-based payment – Good delivery: VCG playment rule – Bad delivery Payment Bidder Bidder Auctioneer 1 2 Bidder i Item delivery November 13, 2015 13/19

5. Trust learning 𝑢 + 𝛽 𝑢 1 − 𝜐 𝑘,𝑠 𝑓 + 𝛽 𝑢 1 − 𝜐 𝑘,𝑠 𝑢 𝑓 = 𝜐 𝑘,𝑠 if 𝑢′ 𝑗,𝑘,𝑙 ≤ 𝑢 𝑗,𝑘,𝑙 = 𝜐 𝑘,𝑠 if 𝑓′ 𝑗,𝑘,𝑙 ≤ 𝑓 𝑗,𝑘,𝑙 𝑢 𝑓 𝜐 𝑘,𝑠+1 𝜐 𝑘,𝑠+1 𝑓 − 𝛾 𝑢 𝜐 𝑗,𝑠 𝑢 − 𝛾 𝑢 𝜐 𝑗,𝑠 e 𝑢 𝜐 𝑗,𝑠 otherwise 𝜐 𝑗,𝑠 otherwise 𝛽 = 𝛾 = 0.01 November 13, 2015 14/19

Experiments  Experiments based on a real business process – One auctioneer outsources tasks to external agents – Consideration of economic cost + delivery time + energy consumption – Greedy bidders – Execution times and energy consumptions based on real agents probability distributions  6 accurate bidders + 6 inaccurate bidders  Each accurate bidder has its own inaccurate twin bidder – Same abilities – Same time and energy distributions November 13, 2015 15/19

Results  The use of trust highly reduces the amount of bad delivered tasks  With agents that always behave equal, Schillo model outperforms the others November 13, 2015 16/19

Results  All bidders misestimate the attributes but good bidders add a security margin ( 1.5 × 𝜏 ) November 13, 2015 17/19

Conclusions  Merge of trust with multi-attribute auctions  Inclusion of trust in the valuation function. This affects: • The winner determination problem • The payment  Flexibility of trust regarding each checkable attribute  Proposal of a trust learning model  Easy to parametrize and adjust the learning curve  It does not present rigidity when faces agents’ behavior changes  Robust against initialization and random misdeliveries November 13, 2015 18/19

T HANKS !!