Optimal Decision Making with CP-nets and PCP-nets

Optimal Decision Making
with CP-nets and PCP-nets
Sibel Adali, Sujoy Sikdar, Lirong Xia
Multi-Issue Voting
{
,
}X{ , }
Main dishes (𝑀)
β€’ 𝑝 issues
β€’ Combinatorial preferences
over decisions on all issues.
Wine (π‘Š)
Goal: Cater to people’s preferences
What is the best decision for all issues?
How to compare two decisions?
β€’ Challenges:
β€’ Preference representation
β€’ Computational
Compact Preference Languages and CP-nets
[Boutilier et al. β€˜04]
β€œI prefer red wine to white wine with my meal,
ceteris paribus, given that meat is served.”
Winners are Undominated
β€’ No other decision is preferred
β€’ Acyclic CP-nets:
β€’ Always exists
β€’ Unique
β€’ Cyclic CP-nets: ???
β€’ Doesn’t always exist
β€’ May not be unique
PCP-nets
[Bigot et al. ’13, Cornelio et al. β€˜13]
Induces CP-net
w/ probability
0.7 × 0.6 × 0.7
PCP-nets are useful
β€’ Uncertain preferences [Bigot et al. ’13, Cornelio et al. β€˜13]
β€’ Dynamic preferences [Cornelio et al. β€˜14]
β€’ Aggregate CP-net profile as a single PCP-net [Cornelio et al. β€˜14]
Previous Work
β€’ Winner determination
β€’ Common assumptions:
β€’ Dependencies are acyclic
β€’ All preferences have the same structure
Quantitative approach to decision making
β€’ Loss Minimization Framework
β€’ # (weakly) dominating decisions
β€’ Optimal decision = Loss minimizing decision
Main messages
β€’ Full generality w.r.t. preferences:
β€’ Cyclic dependencies
β€’ CP-net and PCP-net profiles
β€’ Natural notions of loss
β€’ Generalizes previous work
β€’ New class of voting rules
β€’ And axiomatic properties
Loss of a decision
β€’ Inputs:
β€’ (P)CP-net 𝐢
β€’ Decision 𝑑
β€’ Output:
β€’ Loss 𝐿(𝐢, 𝑑)
Natural loss functions
β€’ 0-1 loss (𝐿0βˆ’1 )
β€’ Is it dominated?
β€’ Most probable optimal decision [Cornelio et al. β€˜13]
β€’ Neighborhood loss (𝐿𝑁 )
β€’ # (weakly) dominating neighbors
β€’ Local Condorcet winner [Conitzer et al. β€˜11]
β€’ Global loss (𝐿𝐺 )
β€’ Total # dominating decisions
𝐿0βˆ’1 (
, )=1
𝐿𝑁 (
, )=2
𝐿𝐺 (
, )=3
Computing the Loss
𝐋𝐨𝐬𝐬 𝐟𝐧.
𝐿0βˆ’1
𝐿𝑁
𝐿𝐺
Acyclic
P (trivial)
coNP-hard
Cyclic
P
P
coNP-hard
Computing the Optimal Decision for CP-nets
Input: CP-net
Output: Optimal decision = Loss minimizing decision
𝐋𝐨𝐬𝐬 𝐟𝐧.
𝐿0βˆ’1
𝐿𝑁
𝐿𝐺
Acyclic
P [Boutilier et al., β€˜04]
Cyclic
NP-complete
P
Optimal Decision for PCP-nets
𝐋𝐨𝐬𝐬 𝐟𝐧.
𝐿0βˆ’1
𝐿𝑁
𝐿𝐺
Acyclic
NP-complete,
P for trees [Cornelio et al., β€˜13]
NP-hard,
P for trees*
Cyclic
NP-complete
[Cornelio et al., β€˜13]
NP-hard
coNP-hard
* Exponential in tree-width
Using a variable-elimination algorithm
A new class of voting rules
β€’ Input: CP-net profile 𝑃
β€’ Output: Decision for every issue
𝐋𝐨𝐬𝐬 𝐟𝐧.
𝐿0βˆ’1
𝐿𝑁
𝐿𝐺
Acyclic
P
Cyclic
NP-complete
NP-complete
P for shared tree dependency structure
coNP-hard
Axiomatic Properties
β€’ Anonymity
β€’ Consistency
β€’ Issue-wise neutrality
β€’ Weak monotonicity
β€’ Satisfied by every rule
Summary and Conclusions
β€’ Quantitative approach to multi-issue voting
β€’ Fully general:
β€’ Cyclic dependencies
β€’ CP-net and PCP-net profiles
β€’ New loss minimization framework
β€’ Natural loss functions
β€’ New class of voting rules
β€’ Identifying tractable sub-cases for optimal outcome of PCP-nets
β€’ Large space of possible loss functions
β€’ Good social choice normative properties
β€’ Computationally tractable