[74] | 1 | import logging
|
---|
| 2 | import time
|
---|
| 3 | from typing import cast
|
---|
| 4 | from operator import itemgetter
|
---|
| 5 | import numpy as np
|
---|
| 6 | from decimal import Decimal
|
---|
| 7 |
|
---|
| 8 | from geniusweb.actions.Accept import Accept
|
---|
| 9 | from geniusweb.actions.Action import Action
|
---|
| 10 | from geniusweb.actions.Offer import Offer
|
---|
| 11 | from geniusweb.bidspace.AllBidsList import AllBidsList
|
---|
| 12 | from geniusweb.inform.ActionDone import ActionDone
|
---|
| 13 | from geniusweb.inform.Finished import Finished
|
---|
| 14 | from geniusweb.inform.Inform import Inform
|
---|
| 15 | from geniusweb.inform.Settings import Settings
|
---|
| 16 | from geniusweb.inform.YourTurn import YourTurn
|
---|
| 17 | from geniusweb.issuevalue.Bid import Bid
|
---|
| 18 | from geniusweb.party.Capabilities import Capabilities
|
---|
| 19 | from geniusweb.party.DefaultParty import DefaultParty
|
---|
| 20 | from geniusweb.profileconnection.ProfileConnectionFactory import (
|
---|
| 21 | ProfileConnectionFactory,
|
---|
| 22 | )
|
---|
| 23 | from geniusweb.progress.ProgressRounds import ProgressRounds
|
---|
| 24 | from .FreqModelWeighted import FreqModelWeighted
|
---|
| 25 | from tudelft_utilities_logging.Reporter import Reporter
|
---|
| 26 |
|
---|
| 27 | """
|
---|
| 28 | BeanBot agent
|
---|
| 29 | """
|
---|
| 30 | class Agent52(DefaultParty):
|
---|
| 31 |
|
---|
| 32 | def __init__(self, reporter: Reporter = None):
|
---|
| 33 | super().__init__(reporter)
|
---|
| 34 | self.getReporter().log(logging.INFO, "party is initialized")
|
---|
| 35 | self._profile = None
|
---|
| 36 | self._last_received_bid: Bid = None
|
---|
| 37 | self._last_received_action = None
|
---|
| 38 | self._opp_model = None
|
---|
| 39 | self._window_size = 10 # last 10 opponent bids are stored window below
|
---|
| 40 | self._opp_bids_window = []
|
---|
| 41 | self._opp_best_bid = None
|
---|
| 42 |
|
---|
| 43 | def notifyChange(self, info: Inform):
|
---|
| 44 | """This is the entry point of all interaction with your agent after is has been initialised.
|
---|
| 45 |
|
---|
| 46 | Args:
|
---|
| 47 | info (Inform): Contains either a request for action or information.
|
---|
| 48 | """
|
---|
| 49 |
|
---|
| 50 | # a Settings message is the first message that will be send to your
|
---|
| 51 | # agent containing all the information about the negotiation session.
|
---|
| 52 | if isinstance(info, Settings):
|
---|
| 53 | self._settings: Settings = cast(Settings, info)
|
---|
| 54 | self._me = self._settings.getID()
|
---|
| 55 |
|
---|
| 56 | # progress towards the deadline has to be tracked manually through the use of the Progress object
|
---|
| 57 | self._progress: ProgressRounds = self._settings.getProgress()
|
---|
| 58 |
|
---|
| 59 | # the profile contains the preferences of the agent over the domain
|
---|
| 60 | self._profile = ProfileConnectionFactory.create(
|
---|
| 61 | info.getProfile().getURI(), self.getReporter()
|
---|
| 62 | )
|
---|
| 63 |
|
---|
| 64 | # Create the weighted frequency model
|
---|
| 65 | self._opp_model = FreqModelWeighted.create().With(self._profile.getProfile().getDomain(), None)
|
---|
| 66 | self._opp_model.__class__ = FreqModelWeighted
|
---|
| 67 |
|
---|
| 68 |
|
---|
| 69 | # Generate sorted (decr.) list of all possible bids with their corresponding utility values
|
---|
| 70 | # Create reservation value after
|
---|
| 71 | profile = self._profile.getProfile()
|
---|
| 72 | allBids = AllBidsList(self._profile.getProfile().getDomain())
|
---|
| 73 | self._bid_utility_tuple = [(bid, profile.getUtility(bid)) for bid in allBids]
|
---|
| 74 | self._bid_utility_tuple.sort(key=itemgetter(1), reverse=True)
|
---|
| 75 |
|
---|
| 76 | # set reservation value to maximum of (0.4, worst bid utility in domain)
|
---|
| 77 | alpha = 0.4
|
---|
| 78 | self._rsv_val = alpha if alpha > self._bid_utility_tuple[-1][1] else self._bid_utility_tuple[-1][1]
|
---|
| 79 |
|
---|
| 80 | # ActionDone is an action send by an opponent (an offer or an accept)
|
---|
| 81 | elif isinstance(info, ActionDone):
|
---|
| 82 | action: Action = cast(ActionDone, info).getAction()
|
---|
| 83 |
|
---|
| 84 | # if it is an offer, set the last received bid
|
---|
| 85 | if isinstance(action, Offer):
|
---|
| 86 | self._last_received_bid = cast(Offer, action).getBid()
|
---|
| 87 | self._last_received_action = action
|
---|
| 88 | # YourTurn notifies you that it is your turn to act
|
---|
| 89 | elif isinstance(info, YourTurn):
|
---|
| 90 | action = self._myTurn()
|
---|
| 91 | if isinstance(self._progress, ProgressRounds):
|
---|
| 92 | self._progress = self._progress.advance()
|
---|
| 93 | self.getConnection().send(action)
|
---|
| 94 |
|
---|
| 95 | # Finished will be send if the negotiation has ended (through agreement or deadline)
|
---|
| 96 | elif isinstance(info, Finished):
|
---|
| 97 | # terminate the agent MUST BE CALLED
|
---|
| 98 | self.terminate()
|
---|
| 99 | else:
|
---|
| 100 | self.getReporter().log(
|
---|
| 101 | logging.WARNING, "Ignoring unknown info " + str(info)
|
---|
| 102 | )
|
---|
| 103 |
|
---|
| 104 | # lets the geniusweb system know what settings this agent can handle
|
---|
| 105 | # leave it as it is for this competition
|
---|
| 106 | def getCapabilities(self) -> Capabilities:
|
---|
| 107 | return Capabilities(
|
---|
| 108 | set(["SAOP"]),
|
---|
| 109 | set(["geniusweb.profile.utilityspace.LinearAdditive"]),
|
---|
| 110 | )
|
---|
| 111 |
|
---|
| 112 | # terminates the agent and its connections
|
---|
| 113 | # leave it as it is for this competition
|
---|
| 114 | def terminate(self):
|
---|
| 115 | self.getReporter().log(logging.INFO, "party is terminating:")
|
---|
| 116 | super().terminate()
|
---|
| 117 | if self._profile is not None:
|
---|
| 118 | self._profile.close()
|
---|
| 119 | self._profile = None
|
---|
| 120 |
|
---|
| 121 |
|
---|
| 122 |
|
---|
| 123 | # give a description of your agent
|
---|
| 124 | def getDescription(self) -> str:
|
---|
| 125 | return "BeanBot implementation for Collaborative AI course"
|
---|
| 126 |
|
---|
| 127 | # execute a turn
|
---|
| 128 | def _myTurn(self):
|
---|
| 129 | # Update the frequency model and the issue weights with the last received bid
|
---|
| 130 | self._opp_model = self._opp_model.WithAction(self._last_received_action, self._progress)
|
---|
| 131 | self._opp_model.__class__ = FreqModelWeighted
|
---|
| 132 | self._opp_model.updateIssueWeights()
|
---|
| 133 |
|
---|
| 134 | # Update the best bid offered by the opponent if the last received bid is better for us
|
---|
| 135 | profile = self._profile.getProfile()
|
---|
| 136 | self._opp_best_bid = self._last_received_bid \
|
---|
| 137 | if self._opp_best_bid is None or profile.getUtility(self._last_received_bid) > profile.getUtility(self._opp_best_bid) \
|
---|
| 138 | else self._opp_best_bid
|
---|
| 139 |
|
---|
| 140 | # Update the bids in the window of last received bids (window has size self._window_size)
|
---|
| 141 | if self._last_received_bid is not None:
|
---|
| 142 | self._opp_bids_window.append(profile.getUtility(self._last_received_bid))
|
---|
| 143 | if len(self._opp_bids_window) > self._window_size:
|
---|
| 144 | self._opp_bids_window.pop(0)
|
---|
| 145 |
|
---|
| 146 | bid = self._findBid()
|
---|
| 147 | action = Offer(self._me, bid)
|
---|
| 148 | # AC_combi check whether we should accept current offer or not
|
---|
| 149 | if self._isGood(self._last_received_bid, bid):
|
---|
| 150 | action = Accept(self._me, self._last_received_bid)
|
---|
| 151 |
|
---|
| 152 | # send the action
|
---|
| 153 | return action
|
---|
| 154 |
|
---|
| 155 |
|
---|
| 156 | """
|
---|
| 157 | AC_combi hybrid acceptance strategy.
|
---|
| 158 | Uses AC_next condition in the first [0, T) fraction of the negotiation.
|
---|
| 159 | Additionally accepts some form of best opponent offer in phase [T, 1] of the negotiation (end phase).
|
---|
| 160 | This can be either if the received utility is better than overall best received bid, best bid in window,
|
---|
| 161 | or average utility in window.
|
---|
| 162 | """
|
---|
| 163 | def _isGood(self, offeredBid: Bid, nextBid: Bid, T=0.9) -> bool:
|
---|
| 164 | if offeredBid is None:
|
---|
| 165 | return False
|
---|
| 166 |
|
---|
| 167 | # progress represents fraction of negotiation that has passed
|
---|
| 168 | profile = self._profile.getProfile()
|
---|
| 169 | progress = self._progress.get(time.time() * 1000)
|
---|
| 170 |
|
---|
| 171 | # AC_next or AC_combi if we are in phase [T, 1]
|
---|
| 172 | accept = profile.getUtility(offeredBid) > profile.getUtility(nextBid) \
|
---|
| 173 | or (progress > T and self._window_max())
|
---|
| 174 | return accept
|
---|
| 175 |
|
---|
| 176 | """
|
---|
| 177 | The following three methods represent the three possible AC_combi methods described above.
|
---|
| 178 | """
|
---|
| 179 | def _window_max(self):
|
---|
| 180 | # check if better than maximum utility in past window
|
---|
| 181 | profile = self._profile.getProfile()
|
---|
| 182 | return profile.getUtility(self._last_received_bid) >= np.max(self._opp_bids_window)
|
---|
| 183 |
|
---|
| 184 | def _window_avg(self):
|
---|
| 185 | # check if better than average utility in past window
|
---|
| 186 | profile = self._profile.getProfile()
|
---|
| 187 | return profile.getUtility(self._last_received_bid) >= np.mean(self._opp_bids_window)
|
---|
| 188 |
|
---|
| 189 | def _overall_max(self):
|
---|
| 190 | # check if better than maximum utility received in entire negotiation
|
---|
| 191 | profile = self._profile.getProfile()
|
---|
| 192 | return profile.getUtility(self._last_received_bid) >= profile.getUtility(self._opp_best_bid)
|
---|
| 193 |
|
---|
| 194 | """
|
---|
| 195 | Implements bidding strategy inspired by the AgreeableAgent2018 (ANAC2018).
|
---|
| 196 | Take all bids higher than target utility and pick random one based on opponent preferences to send to opponent.
|
---|
| 197 | """
|
---|
| 198 | def _findBid(self) -> Bid:
|
---|
| 199 | # e value determines concession rate by influencing the shape of the target utility curve
|
---|
| 200 | e = 0.3
|
---|
| 201 | max_util = self._bid_utility_tuple[0][1]
|
---|
| 202 | target_util = self._getUtilityGoal(self._progress.get(time.time() * 1000), e, Decimal(self._rsv_val), max_util)
|
---|
| 203 | # Allow for some additional (10% of target utility) randomness in the possible bids to send to opponent
|
---|
| 204 | # in the first half of the negotiation. Otherwise, will send mostly the same bid constantly at first.
|
---|
| 205 | if self._progress.get(time.time() * 1000) < 0.5:
|
---|
| 206 | target_util = target_util - Decimal(np.random.uniform(0, 0.1 * float(target_util)))
|
---|
| 207 |
|
---|
| 208 | candidates = []
|
---|
| 209 | opp_utilities = []
|
---|
| 210 | # Find all bids above target utility and store along with the associated opponent utilities of the bids
|
---|
| 211 | for items in self._bid_utility_tuple:
|
---|
| 212 | if items[1] < target_util:
|
---|
| 213 | break
|
---|
| 214 | candidates.append(items[0])
|
---|
| 215 | opp_utilities.append(float(self._opp_model.getUtility(items[0])))
|
---|
| 216 |
|
---|
| 217 | # apply roulette wheel selection to the bids to choose one using exponential fitness function
|
---|
| 218 | return self._roulette_selection(candidates, opp_utilities, self._fitness_exp)
|
---|
| 219 |
|
---|
| 220 | """
|
---|
| 221 | Roulette wheel selection to select a random bid to send.
|
---|
| 222 | Scale opponent utilities to [0, 1], apply fitness function to it and use fitness values as probabilities
|
---|
| 223 | for choosing each bid.
|
---|
| 224 | """
|
---|
| 225 | def _roulette_selection(self, candidates, utilities, fitness_func, eps=0.0001):
|
---|
| 226 | normalised_utils = np.array(utilities)
|
---|
| 227 |
|
---|
| 228 | # if same utilities, choose random bid, otherwise continue scaling to [0, 1]
|
---|
| 229 | if np.max(normalised_utils) - np.min(normalised_utils) < eps:
|
---|
| 230 | return np.random.choice(candidates)
|
---|
| 231 |
|
---|
| 232 | # scale utilities to [0, 1]
|
---|
| 233 | normalised_utils = (normalised_utils - np.min(normalised_utils)) / (np.max(normalised_utils) - np.min(normalised_utils))
|
---|
| 234 | # apply fitness function and divide by total sum in order to obtain probability values for each bid
|
---|
| 235 | fitnesses = fitness_func(normalised_utils)
|
---|
| 236 | fitnesses /= np.sum(fitnesses)
|
---|
| 237 | # return random bid using fitnesses as weights, or completely random bid if something went wrong in fitnesses
|
---|
| 238 | return np.random.choice(candidates, p=fitnesses) if not np.isnan(fitnesses).any() else np.random.choice(candidates)
|
---|
| 239 |
|
---|
| 240 | """
|
---|
| 241 | Next two functions allow for two different shapes for the fitness transformation of the utilities
|
---|
| 242 | """
|
---|
| 243 | def _fitness_linear(self, normalised_utils):
|
---|
| 244 | return 0.3 + 0.7 * normalised_utils
|
---|
| 245 |
|
---|
| 246 | def _fitness_exp(self, normalised_utils, alpha=3):
|
---|
| 247 | return np.exp(-alpha*(1-normalised_utils))
|
---|
| 248 |
|
---|
| 249 | """
|
---|
| 250 | Same function as used by time_dependent_agent to determine the curve of utility target line.
|
---|
| 251 | """
|
---|
| 252 | def _getUtilityGoal(
|
---|
| 253 | self, t: float, e: float, minUtil: Decimal, maxUtil: Decimal
|
---|
| 254 | ) -> Decimal:
|
---|
| 255 | """
|
---|
| 256 | @param t the time in [0,1] where 0 means start of nego and 1 the
|
---|
| 257 | end of nego (absolute time/round limit)
|
---|
| 258 | @param e the e value that determinses how fast the party makes
|
---|
| 259 | concessions with time. Typically around 1. 0 means no
|
---|
| 260 | concession, 1 linear concession, >1 faster than linear
|
---|
| 261 | concession.
|
---|
| 262 | @param minUtil the minimum utility possible in our profile
|
---|
| 263 | @param maxUtil the maximum utility possible in our profile
|
---|
| 264 | @return the utility goal for this time and e value
|
---|
| 265 | """
|
---|
| 266 |
|
---|
| 267 | ft1 = Decimal(1)
|
---|
| 268 | if e != 0:
|
---|
| 269 | ft1 = round(Decimal(1 - pow(t, 1 / e)), 6) # defaults ROUND_HALF_UP
|
---|
| 270 | return max(min((minUtil + (maxUtil - minUtil) * ft1), maxUtil), minUtil)
|
---|