def compute_seq_thiele_methods(profile, committeesize, scorefct_str):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
comm_scores = {(): 0}
# build committees starting with the empty set
for _ in range(0, committeesize):
comm_scores_next = {}
for committee, score in comm_scores.items():
# marginal utility gained by adding candidate to the committee
additional_score_cand = sf.additional_thiele_scores(
profile, committee, scorefct)
for c in range(profile.num_cand):
if additional_score_cand[c] >= max(additional_score_cand):
next_comm = tuple(sorted(committee + (c, )))
comm_scores_next[next_comm] = (comm_scores[committee] +
additional_score_cand[c])
# remove suboptimal committees
comm_scores = {}
cutoff = max(comm_scores_next.values())
for com, score in comm_scores_next.items():
if score >= cutoff:
comm_scores[com] = score
return sort_committees(list(comm_scores.keys()))
def compute_monroe_bruteforce(profile,
committeesize,
resolute=False,
flowbased=True):
"""Returns the list of winning committees via brute-force Monroe's rule"""
enough_approved_candidates(profile, committeesize)
if not profile.has_unit_weights():
raise Exception("Monroe is only defined for unit weights (weight=1)")
if profile.totalweight() % committeesize != 0 or flowbased:
monroescore = sf.monroescore_flowbased
else:
monroescore = sf.monroescore_matching
opt_committees = []
opt_monroescore = -1
for comm in combinations(list(range(profile.num_cand)), committeesize):
score = monroescore(profile, comm)
if score > opt_monroescore:
opt_committees = [comm]
opt_monroescore = score
elif monroescore(profile, comm) == opt_monroescore:
opt_committees.append(comm)
opt_committees = sort_committees(opt_committees)
if resolute:
return [opt_committees[0]]
else:
return opt_committees
def compute_av(profile, committeesize, resolute=False, sav=False):
"""Returns the list of winning committees according to Approval Voting"""
enough_approved_candidates(profile, committeesize)
appr_scores = [0] * profile.num_cand
for pref in profile.preferences:
for cand in pref.approved:
if sav:
# Satisfaction Approval Voting
appr_scores[cand] += Fraction(pref.weight, len(pref.approved))
else:
# (Classic) Approval Voting
appr_scores[cand] += pref.weight
# smallest score to be in the committee
cutoff = sorted(appr_scores)[-committeesize]
certain_cand = [
c for c in range(profile.num_cand) if appr_scores[c] > cutoff
]
possible_cand = [
c for c in range(profile.num_cand) if appr_scores[c] == cutoff
]
missing = committeesize - len(certain_cand)
if resolute:
return sort_committees([(certain_cand + possible_cand[:missing])])
else:
return sort_committees([
(certain_cand + list(selection))
for selection in combinations(possible_cand, missing)
])
def compute_seqphragmen(profile, committeesize, resolute=False):
"""Returns the list of winning committees
according to sequential Phragmen"""
enough_approved_candidates(profile, committeesize)
load = {v: 0 for v in profile.preferences}
comm_loads = {(): load}
approvers_weight = {}
for c in range(profile.num_cand):
approvers_weight[c] = sum(v.weight for v in profile.preferences
if c in v.approved)
# build committees starting with the empty set
for _ in range(0, committeesize):
comm_loads_next = {}
for committee, load in comm_loads.items():
approvers_load = {}
for c in range(profile.num_cand):
approvers_load[c] = sum(v.weight * load[v]
for v in profile.preferences
if c in v.approved)
new_maxload = [
Fraction(approvers_load[c] + 1, approvers_weight[c])
if approvers_weight[c] > 0 else committeesize + 1
for c in range(profile.num_cand)
]
for c in range(profile.num_cand):
if c in committee:
new_maxload[c] = sys.maxsize
for c in range(profile.num_cand):
if new_maxload[c] <= min(new_maxload):
new_load = {}
for v in profile.preferences:
if c in v.approved:
new_load[v] = new_maxload[c]
else:
new_load[v] = load[v]
comm_loads_next[tuple(sorted(committee +
(c, )))] = new_load
# remove suboptimal committees
comm_loads = {}
cutoff = min([max(load.values()) for load in comm_loads_next.values()])
for com, load in comm_loads_next.items():
if max(load.values()) <= cutoff:
comm_loads[com] = load
if resolute:
committees = sort_committees(list(comm_loads.keys()))
comm = tuple(committees[0])
comm_loads = {comm: comm_loads[comm]}
committees = sort_committees(list(comm_loads.keys()))
if resolute:
return [committees[0]]
else:
return committees
def compute_revseq_thiele_methods_resolute(profile, committeesize,
scorefct_str):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
committee = set(range(profile.num_cand))
for _ in range(0, profile.num_cand - committeesize):
cands_to_remove, _ = __least_relevant_cands(profile, committee,
scorefct)
committee.remove(cands_to_remove[0])
return [sorted(list(committee))]
def compute_seq_thiele_resolute(profile, committeesize, scorefct_str):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
committee = []
# build committees starting with the empty set
for _ in range(0, committeesize):
additional_score_cand = sf.additional_thiele_scores(
profile, committee, scorefct)
next_cand = additional_score_cand.index(max(additional_score_cand))
committee.append(next_cand)
return [sorted(committee)]
def compute_greedy_monroe(profile, committeesize):
""""Returns the winning committee of the greedy monroe.
Always selects the candidate with the highest approval.
Always removes the first n/k (rounding depends) voters that approve
with the selected candidate. (voter sorted by their rankings)
"""
enough_approved_candidates(profile, committeesize)
if not profile.has_unit_weights():
raise Exception("Greedy Monroe is only defined for unit weights" +
" (weight=1)")
v = list(enumerate(list(profile.preferences)))
# list of tuples (nr, Preferences)
# sorted by sorted approved list of preferences
voters = sorted(v, key=lambda p: sorted(p[1].approved))
n = len(voters) # number of voters
cands = set(range(profile.num_cand))
not_s, committee = (voters, set()) # not_s .. not satisfied voters
for t in range(1, committeesize + 1):
remaining_cands = cands - committee
approval = {c: 0 for c in remaining_cands}
for nr, voter in not_s:
for c in voter.approved:
if c in remaining_cands:
approval[c] += 1
max_approval = max(approval.values())
winner = [c for c in remaining_cands if approval[c] == max_approval][0]
# round how many are removed, either up or down
if t <= n - committeesize * math.floor(n / committeesize):
to_remove = math.ceil(float(n) / committeesize)
else:
to_remove = math.floor(n / committeesize)
# not more than the voters that approve
# the candidate can be removed
to_remove = min(max_approval, to_remove)
next_voters = []
for nr, voter in not_s:
if to_remove > 0 and winner in voter.approved:
to_remove -= 1
else:
next_voters.append((nr, voter))
not_s = next_voters
committee.add(winner)
return sort_committees([committee])
def compute_thiele_methods_branchandbound(profile,
committeesize,
scorefct_str,
resolute=False):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
best_committees = []
init_com = compute_seq_thiele_resolute(profile, committeesize,
scorefct_str)
best_score = sf.thiele_score(profile, init_com[0], scorefct_str)
part_coms = [[]]
while part_coms:
part_com = part_coms.pop(0)
# potential committee, check if at least as good
# as previous best committee
if len(part_com) == committeesize:
score = sf.thiele_score(profile, part_com, scorefct_str)
if score == best_score:
best_committees.append(part_com)
elif score > best_score:
best_committees = [part_com]
best_score = score
else:
if len(part_com) > 0:
largest_cand = part_com[-1]
else:
largest_cand = -1
missing = committeesize - len(part_com)
marg_util_cand = sf.additional_thiele_scores(
profile, part_com, scorefct)
upper_bound = (
sum(sorted(marg_util_cand[largest_cand + 1:])[-missing:]) +
sf.thiele_score(profile, part_com, scorefct_str))
if upper_bound >= best_score:
for c in range(largest_cand + 1,
profile.num_cand - missing + 1):
part_coms.insert(0, part_com + [c])
committees = sort_committees(best_committees)
if resolute:
return [committees[0]]
else:
return committees
def compute_minimaxav(profile, committeesize, ilp=True, resolute=False):
"""Returns the list of winning committees according to Minimax AV"""
if ilp:
return compute_minimaxav_ilp(profile, committeesize, resolute)
def hamming(a, b, elements):
diffs = 0
for x in elements:
if (x in a and x not in b) or (x in b and x not in a):
diffs += 1
return diffs
def mavscore(committee, profile):
score = 0
for vote in profile.preferences:
hamdistance = hamming(vote.approved, committee,
list(range(profile.num_cand)))
if hamdistance > score:
score = hamdistance
return score
enough_approved_candidates(profile, committeesize)
opt_committees = []
opt_mavscore = profile.num_cand + 1
for comm in combinations(list(range(profile.num_cand)), committeesize):
score = mavscore(comm, profile)
if score < opt_mavscore:
opt_committees = [comm]
opt_mavscore = score
elif mavscore(comm, profile) == opt_mavscore:
opt_committees.append(comm)
opt_committees = sort_committees(opt_committees)
if resolute:
return [opt_committees[0]]
else:
return sort_committees(opt_committees)
def compute_revseq_thiele_methods(profile, committeesize, scorefct_str):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
allcandcomm = tuple(range(profile.num_cand))
comm_scores = {
allcandcomm: sf.thiele_score(profile, allcandcomm, scorefct_str)
}
for _ in range(0, profile.num_cand - committeesize):
comm_scores_next = {}
for committee, score in comm_scores.items():
cands_to_remove, score_reduction = \
__least_relevant_cands(profile, committee, scorefct)
for c in cands_to_remove:
next_comm = tuple(set(committee) - set([c]))
comm_scores_next[next_comm] = score - score_reduction
# remove suboptimal committees
comm_scores = {}
cutoff = max(comm_scores_next.values())
for com, score in comm_scores_next.items():
if score >= cutoff:
comm_scores[com] = score
return sort_committees(list(comm_scores.keys()))
def compute_monroe_ilp(profile, committeesize, resolute):
enough_approved_candidates(profile, committeesize)
# Monroe is only defined for unit weights
if not profile.has_unit_weights():
raise Exception("Monroe is only defined for unit weights (weight=1)")
num_voters = len(profile.preferences)
cands = list(range(profile.num_cand))
# Alternative: split voters -> generate new profile with all weights = 1
# prof2 = Profile(profile.num_cand)
# for v in profile:
# for _ in range(v.weight):
# prof2.add_preference(DichotomousPreference(v.approved,
# profile.num_cand))
# total_weight = profile.voters_num()
m = gb.Model()
# optimization goal: variable "satisfaction"
satisfaction = m.addVar(vtype=gb.GRB.INTEGER, name="satisfaction")
# a list of committee members
in_committee = m.addVars(profile.num_cand,
vtype=gb.GRB.BINARY,
name="in_comm")
m.addConstr(gb.quicksum(in_committee[c] for c in cands) == committeesize)
# a partition of voters into committeesize many sets
partition = m.addVars(profile.num_cand,
len(profile.preferences),
vtype=gb.GRB.INTEGER,
lb=0,
name="partition")
for i in range(len(profile.preferences)):
# every voter has to be part of a voter partition set
m.addConstr(
gb.quicksum(partition[(j, i)]
for j in cands) == profile.preferences[i].weight)
for i in cands:
# every voter set in the partition has to contain
# at least (num_voters // committeesize) candidates
m.addConstr(
gb.quicksum(partition[(i, j)]
for j in range(len(profile.preferences))) >=
(num_voters // committeesize - num_voters * (1 - in_committee[i])))
# every voter set in the partition has to contain
# at most ceil(num_voters/committeesize) candidates
m.addConstr(
gb.quicksum(partition[(i, j)]
for j in range(len(profile.preferences))) <=
(num_voters // committeesize + bool(num_voters % committeesize) +
num_voters * (1 - in_committee[i])))
# if in_committee[i] = 0 then partition[(i,j) = 0
m.addConstr(
gb.quicksum(
partition[(i, j)]
for j in range(len(profile.preferences))) <= num_voters *
in_committee[i])
m.update()
# constraint for objective variable "satisfaction"
m.addConstr(
gb.quicksum(partition[(i, j)] * (i in profile.preferences[j].approved)
for j in range(len(profile.preferences))
for i in cands) >= satisfaction)
# optimization objective
m.setObjective(satisfaction, gb.GRB.MAXIMIZE)
# m.setParam('OutputFlag', False)
m.setParam('OutputFlag', False)
if resolute:
m.setParam('PoolSearchMode', 0)
else:
# output all optimal committees
m.setParam('PoolSearchMode', 2)
# abort after (roughly) 100 optimal solutions
m.setParam('PoolSolutions', 100)
# ignore suboptimal committees
m.setParam('PoolGap', 0)
m.optimize()
if m.Status != 2:
print "Warning (Monroe): solutions may be incomplete or not optimal."
print "(Gurobi return code", m.Status, ")"
# extract committees from model
committees = []
if resolute:
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
else:
for sol in range(m.SolCount):
m.setParam('SolutionNumber', sol)
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
# if len(committees)>10:
# print "Warning (Monroe): more than 10 committees found;",
# print "returning first 10"
# committees = committees[:10]
committees = sort_committees(committees)
return committees
def compute_thiele_methods_ilp(profile,
committeesize,
scorefct_str,
resolute=False):
enough_approved_candidates(profile, committeesize)
scorefct = sf.get_scorefct(scorefct_str, committeesize)
m = gb.Model()
cands = list(range(profile.num_cand))
# a binary variable indicating whether c is in the committee
in_committee = m.addVars(profile.num_cand,
vtype=gb.GRB.BINARY,
name="in_comm")
# a (intended binary) variable indicating
# whether v approves at least l candidates in the committee
utility = {}
for v in profile.preferences:
for l in range(1, committeesize + 1):
utility[(v, l)] = m.addVar(ub=1.0)
# constraint: the committee has the required size
m.addConstr(gb.quicksum(in_committee[c] for c in cands) == committeesize)
# constraint: utilities are consistent with actual committee
for v in profile.preferences:
m.addConstr(
gb.quicksum(utility[v, l]
for l in range(1, committeesize + 1)) == gb.quicksum(
in_committee[c] for c in v.approved))
# objective: the PAV score of the committee
m.setObjective(
gb.quicksum(
float(scorefct(l)) * v.weight * utility[(v, l)]
for v in profile.preferences
for l in range(1, committeesize + 1)), gb.GRB.MAXIMIZE)
m.setParam('OutputFlag', False)
if resolute:
m.setParam('PoolSearchMode', 0)
else:
# output all optimal committees
m.setParam('PoolSearchMode', 2)
# abort after (roughly) 100 optimal solutions
m.setParam('PoolSolutions', 100)
# ignore suboptimal committees
m.setParam('PoolGap', 0)
m.optimize()
if m.Status != 2:
print "Warning (" + scorefct_str + "):",
print "solutions may be incomplete or not optimal."
print "(Gurobi return code", m.Status, ")"
# extract committees from model
committees = []
if resolute:
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
else:
for sol in range(m.SolCount):
m.setParam('SolutionNumber', sol)
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
committees = sort_committees(committees)
return committees
def compute_minimaxav_ilp(profile, committeesize, resolute=False):
enough_approved_candidates(profile, committeesize)
voters = profile.preferences
num_voters = len(voters)
cands = list(range(profile.num_cand))
m = gb.Model()
# optimization goal: variable "sum_difference"
max_hamdistance = m.addVar(vtype=gb.GRB.INTEGER, name="max_hamdistance")
# a list of committee members
in_committee = m.addVars(profile.num_cand,
vtype=gb.GRB.BINARY,
name="in_comm")
m.addConstr(gb.quicksum(in_committee[c] for c in cands) == committeesize)
# the single differences between the committee and the voters
difference = m.addVars(profile.num_cand,
num_voters,
vtype=gb.GRB.INTEGER,
name="diff")
for i in cands:
for j in range(num_voters):
if i in voters[j].approved:
# constraint for the case that the candidate is approved
m.addConstr(difference[i, j] == 1 - in_committee[i])
else:
# constraint for the case that the candidate isn't approved
m.addConstr(difference[i, j] == in_committee[i])
for j in range(num_voters):
# maximum hamming distance is greater of equal than any individual one
m.addConstr(max_hamdistance >= gb.quicksum(difference[i, j]
for i in cands))
# optimization objective
m.setObjective(max_hamdistance, gb.GRB.MINIMIZE)
# m.setParam('OutputFlag', False)
m.setParam('OutputFlag', False)
if resolute:
m.setParam('PoolSearchMode', 0)
else:
# output all optimal committees
m.setParam('PoolSearchMode', 2)
# abort after (roughly) 100 optimal solutions
m.setParam('PoolSolutions', 1000)
# ignore suboptimal committees
m.setParam('PoolGap', 0)
m.optimize()
if m.Status != 2:
print(
"Warning (Minimax AV): solutions may be incomplete or not optimal."
)
print("(Gurobi return code", m.Status, ")")
# extract committees from model
committees = []
if resolute:
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
else:
for sol in range(m.SolCount):
m.setParam('SolutionNumber', sol)
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
committees = sort_committees(committees)
return committees
def compute_optphragmen_ilp(profile, committeesize, resolute=False):
enough_approved_candidates(profile, committeesize)
cands = list(range(profile.num_cand))
m = gb.Model()
m.setParam('OutputFlag', False)
# a binary variable indicating whether c is in the committee
in_committee = m.addVars(profile.num_cand,
vtype=gb.GRB.BINARY,
name="in_comm")
load = {}
for c in cands:
for v in profile.preferences:
load[(v, c)] = m.addVar(ub=1.0, lb=0.0)
# constraint: the committee has the required size
m.addConstr(gb.quicksum(in_committee[c] for c in cands) == committeesize)
for c in cands:
for v in profile.preferences:
if c not in v.approved:
m.addConstr(load[(v, c)] == 0)
# a candidate's load is distributed among his approvers
for c in cands:
m.addConstr(
gb.quicksum(v.weight * load[(v, c)] for v in profile.preferences
if c in cands) == in_committee[c])
loadbound = m.addVar(name="loadbound")
for v in profile.preferences:
m.addConstr(gb.quicksum(load[(v, c)] for c in v.approved) <= loadbound)
m.setObjective(loadbound, gb.GRB.MINIMIZE)
m.setParam('OutputFlag', False)
if resolute:
m.setParam('PoolSearchMode', 0)
else:
# output all optimal committees
m.setParam('PoolSearchMode', 2)
# abort after (roughly) 100 optimal solutions
m.setParam('PoolSolutions', 100)
# ignore suboptimal committees
m.setParam('PoolGap', 0)
m.optimize()
if m.Status != 2:
print("Warning (opt-Phragmen): solutions may be " +
"incomplete or not optimal.")
print("(Gurobi return code", m.Status, ")")
# extract committees from model
committees = []
if resolute:
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
else:
for sol in range(m.SolCount):
m.setParam('SolutionNumber', sol)
committees.append([c for c in cands if in_committee[c].Xn >= 0.99])
committees = sort_committees(committees)
return committees
def compute_phragmen_enestroem(profile, committeesize, resolute=False):
""""Returns the winning committees with
Phragmen's first method (Enestroem's method) –
STV with unordered ballots
In every step the candidate with the highest combined budget of
their supporters gets into a committee.
For equal voting power multiple committees are computed.
Method from:
https://arxiv.org/pdf/1611.08826.pdf (18.5, Page 59)
"""
enough_approved_candidates(profile, committeesize)
num_voters = len(profile.preferences)
start_budget = {
v: Fraction(profile.preferences[v].weight)
for v in range(num_voters)
}
price = Fraction(sum(start_budget.values()), committeesize)
cands = range(profile.num_cand)
committees = [(start_budget, set())]
for i in range(committeesize):
# here the committees with i+1 candidates are
# stored (together with budget)
next_committees = []
# loop in case multiple possible committees
# with i filled candidates
for committee in committees:
budget, comm = committee
curr_cands = set(cands) - comm
support = {c: 0 for c in curr_cands}
for nr, pref in enumerate(profile.preferences):
voting_power = budget[nr]
if voting_power <= 0:
continue
for cand in pref.approved:
if cand in curr_cands:
support[cand] += voting_power
max_support = max(support.values())
winners = [c for c, s in support.items() if s == max_support]
for cand in winners:
b = dict(budget) # new copy of budget
if max_support > price: # supporters can afford it
# (voting_power - price) / voting_power
multiplier = Fraction(max_support - price, max_support)
else: # set supporters to 0
multiplier = 0
for nr, pref in enumerate(profile.preferences):
if cand in pref.approved:
b[nr] *= multiplier
c = comm.union([cand]) # new committee with candidate
next_committees.append((b, c))
if resolute: # only one is requested
if len(next_committees) > 0:
committees = [next_committees[0]]
else: # should not happen
committees = []
raise Exception(
"phragmen enestroem failed to find " +
"next candidate for", committees)
else:
committees = next_committees
committees = [comm for b, comm in committees]
committees = sort_committees(committees)
if resolute:
if len(committees) > 0:
return [committees[0]]
else:
return []
else:
return committees
def compute_rule_x(profile, committeesize, resolute=False):
"""Returns the list of winning candidates according to rule x.
But rule x does stop if not enough budget is there to finance a
candidate. As this is not optimal the committee is filled with the
candidates that have the most remaining budget as support.
Rule from:
https://arxiv.org/pdf/1911.11747.pdf (Page 7)"""
enough_approved_candidates(profile, committeesize)
if not profile.has_unit_weights():
raise Exception("Rule X is only defined \
for unit weights (weight=1)")
num_voters = len(profile.preferences)
price = Fraction(num_voters, committeesize)
start_budget = {v: Fraction(1, 1) for v in range(num_voters)}
cands = range(profile.num_cand)
committees = [(start_budget, set())]
final_committees = []
for _ in range(committeesize):
next_committees = []
for committee in committees:
budget = committee[0]
q_affordability = {}
curr_cands = set(cands) - committee[1]
for c in curr_cands:
approved_by = set()
for v, vote in enumerate(profile.preferences):
if c in vote.approved and budget[v] > 0.0:
approved_by.add(v)
too_poor = set()
already_available = Fraction(0)
rich = set(approved_by)
q = 0.0
while already_available < price and q == 0.0 and len(rich) > 0:
fair_split = Fraction(price - already_available, len(rich))
still_rich = set()
for v in rich:
if budget[v] <= fair_split:
too_poor.add(v)
already_available += budget[v]
else:
still_rich.add(v)
if len(still_rich) == len(rich):
q = fair_split
q_affordability[c] = q
elif already_available == price:
q = fair_split
q_affordability[c] = q
else:
rich = still_rich
if len(q_affordability) > 0:
min_q = min(q_affordability.values())
cheapest_split = [
c for c in q_affordability if q_affordability[c] == min_q
]
for c in cheapest_split:
b = dict(committee[0])
for v, vote in enumerate(profile.preferences):
if c in vote.approved:
b[v] -= min(budget[v], min_q)
comm = set(committee[1])
comm.add(c)
next_committees.append((b, comm))
else: # no affordable candidate remains
comms = fill_remaining_committee(committee, curr_cands,
committeesize, profile)
# after filling the remaining spots these committees
# have size committeesize
for b, comm in comms:
final_committees.append(comm)
if resolute:
if len(next_committees) > 0:
committees = [next_committees[0]]
else:
committees = []
else:
committees = next_committees
# The committees that could be fully filled with Rule X:
for b, comm in committees: # budget and committee
final_committees.append(comm)
committees = sort_committees(final_committees)
if resolute:
if len(committees) > 0:
return [committees[0]]
else:
return []
else:
return committees