source: src/main/java/agents/rlboa/QlearningStrategy.java@ 346

package agents.rlboa;

import genius.core.StrategyParameters;
import genius.core.bidding.BidDetails;
import genius.core.boaframework.NegotiationSession;
import genius.core.boaframework.OfferingStrategy;
import genius.core.boaframework.OpponentModel;
import genius.core.boaframework.OutcomeSpace;
import genius.core.boaframework.SortedOutcomeSpace;
import genius.core.misc.Range;

import java.util.ArrayList;
import java.util.Collections;
import java.util.HashMap;
import java.util.List;
import java.util.Random;

public class QlearningStrategy extends OfferingStrategy {

        protected HashMap<Integer, ArrayList<Double>> qTable;
        protected ArrayList<Integer> actions = new ArrayList<Integer>();
        protected int bins;
        protected Double eps;
        protected Double alpha;
        protected Double gamma;
        protected AbstractState state;
        protected String mode;
        protected int timeBins;
        protected Range minMaxBin;

        public QlearningStrategy(NegotiationSession negotiationSession, OpponentModel opponentModel) {
                super.init(negotiationSession, null);
                this.opponentModel = opponentModel;
                this.endNegotiation = false;
                this.state = State.INITIAL;
                                
                OutcomeSpace outcomeSpace = new SortedOutcomeSpace(negotiationSession.getUtilitySpace());
                this.negotiationSession.setOutcomeSpace(outcomeSpace);
        }

        public ArrayList<Integer> getActions() {
                return this.actions;
        }

        /**
         * @return int representing the last action taken by the strategy
         * @throws IndexOutOfBoundsException if called before any action has been performed
         */
        public int getLastAction() throws IndexOutOfBoundsException {
                return this.actions.get(this.actions.size() - 1);
        }
        
        public void setMinMaxBin(Range minMaxBin) {
                this.minMaxBin = minMaxBin;
        }
        
        public Range getMinMaxBin() {
                return this.minMaxBin;
        }

        protected void initQTable() {
                this.qTable = new HashMap<Integer, ArrayList<Double>>();

                // Initial state has different action space
                this.qTable.putIfAbsent(this.state.hash(), new ArrayList<Double>(Collections.nCopies(this.state.getActionSize(), 0.0)));
        }

        public void initQtable(HashMap<Integer, ArrayList<Double>> qTable) {
                if (qTable != null) {
                        this.qTable = qTable;
                }
                else {
                        this.initQTable();
                }
        }

        @Override
        public BidDetails determineOpeningBid() {
                // Open the negotiation with a free bid (one of N bins)
                int targetBin = this.determineOpeningBin();
                return this.pickBidInBin(targetBin);
        }

        @Override
        public BidDetails determineNextBid() {
                // HACK(?) this QlearningStrategy works for all states that represent the world in bins,
                // so we needed a way to recognize these. Therefore the interface BinnedRepresentation
                int targetBin = this.determineTargetBin(((BinnedRepresentation) this.state).getMyBin());
                return this.pickBidInBin(targetBin);
        }

        @Override
        public String getName() {
                return "Q-Offering";
        }

        /**
         * Check if the bid falls inside the lower and upper bounds
         * @param lower lower bound of utility (inclusive)
         * @param upper upper bound of utility (exclusive)
         * @param bidDetails bid to check (has util and time)
         * @return boolean
         */
        private boolean isInBin(double lower, double upper, BidDetails bidDetails) {
                double myUtil = bidDetails.getMyUndiscountedUtil();
                return myUtil < upper && myUtil >= lower;
        }

        /**
         * Make the opponent model select a bid that is in the provided target bin
         * @param targetBin index of the bin in which to pick a bid
         * @return BidDetails of the selected bid
         */
        protected BidDetails pickBidInBin(int targetBin) {

                double lowerBound = targetBin * this.getBinSize();
                double upperBound = lowerBound + this.getBinSize();

                // getBidsInRange behaves weirdly and returns bids that are outise of there range (false positives)
                List<BidDetails> bidsInRange = this.negotiationSession.getOutcomeSpace().getBidsinRange(new Range(lowerBound, upperBound));
                bidsInRange.removeIf( bid -> !this.isInBin(lowerBound, upperBound, bid) );

                // If there are no bids possible within this bin, recursively choose another bin by the following logic: 
                // if you conceded this round, concede further, etc.
                if (bidsInRange.isEmpty()) {

                        Random random = new Random();
                        int newBin = 0;
                        int direction = -1;
                        
                        // Check if this is the opening action or not; if it is we just pick randomly
                        if (this.actions.size() > 1) {
                                direction = this.actions.get(this.actions.size() - 1);
                        } else {
                                newBin = random.nextInt(this.bins);
                        }

                        // conceded last time
                        if (direction == 0) {
                                newBin = determineTargetBin(targetBin - 1);
                        }
                        
                        // retracted last time
                        if (direction == 1) {
                                newBin = determineTargetBin(targetBin + 1);
                        }
                        
                        // stayed last time
                        if (direction == 2) {
                                int randomUpOrDown = random.nextBoolean() ? 1 : -1;
                                newBin = determineTargetBin(targetBin + randomUpOrDown);
                        }
                        
                        return this.pickBidInBin(newBin);
                }
                
                
                return this.maxBidForOpponent(bidsInRange);
        }

        /**
         * This is the general action function for the RL-agent. We determine a bin by either
         * moving up (retracting offer), doing nothing or moving down (conceding offer).
         * @param currentBin
         * @return
         */
        protected int determineTargetBin(int currentBin) {
                int targetBin = currentBin;
                ArrayList<Double> defaultActionValues = new ArrayList<Double>(Collections.nCopies(this.state.getActionSize(), 0.0));

                List<Double> qValues = this.qTable.getOrDefault(this.state.hash(), defaultActionValues);
                int action = this.epsilonGreedy(qValues);
                this.actions.add(action);

                // Apply action current bin (ie. move up, down or stay)
                switch (action) {
                        case 0: targetBin--;
                        break;
                        case 1: targetBin++;
                        break;
                        case 2: break;
                }

                // Can't go outside of the range of relevant bins.
                targetBin = Math.min(targetBin, (int) this.minMaxBin.getUpperbound());
                targetBin = Math.max(targetBin, (int) this.minMaxBin.getLowerbound());

                return targetBin;
        }

        protected int determineOpeningBin() {
                ArrayList<Double> defaultInitialActionValues = new ArrayList<Double>(Collections.nCopies(this.state.getActionSize(), 0.0));
                List<Double> qValues = this.qTable.getOrDefault(this.state.hash(), defaultInitialActionValues);
                int action = this.epsilonGreedy(qValues);
                this.actions.add(action);

                return action;
        }

        /**
         * @param list List of doubles
         * @return The index of the highest value in the list
         */
        protected int indifferentArgMax(List<Double> list) {
                double maximum = Collections.max(list);

                List<Integer> maximaIdxs = new ArrayList<Integer>();

                // collect indices of all occurrences of maximum
                for (int i = 0; i < list.size(); i++) {
                        if (list.get(i) == maximum) {
                                maximaIdxs.add(i);
                        }
                }

                // pick a random index from the list (this is the indifferent part)
                Random rnd = new Random();
                int choice = rnd.nextInt(maximaIdxs.size());

                return maximaIdxs.get(choice);
        }

        protected int epsilonGreedy(List<Double> qValues) {
                int action;

                // With probability epsilon, pick a random action (epsilon greedy)
                if (Math.random() < this.eps && this.isTraining()) {
                        Random random = new Random();
                        action = random.nextInt(qValues.size());
                }
                else {
                        action = this.indifferentArgMax(qValues);
                }

                return action;
        }

        /**
         * @return The number of bins in which the each utility axis is divided
         */
        int getNBins() {
                return this.bins;
        }

        /**
         * @return The width of the bins in which the each utility axis is divided
         */
        protected double getBinSize() {
                return 1.0 / this.getNBins();
        }

        /**
         * Setter for the state property
         * @param state new {@link State}
         *
         */
        protected void setState(State state) {
                this.state = state;
        }

        /**
         * Getter for the state property
         * @return
         */
        protected AbstractState getState() {
                return this.state;
        }

        /**
         * Determine the bid with the highest expected utility for the opponent from a list of bids
         * @param bids
         * @return BidDetails with representing the maximum bid
         */
        protected BidDetails maxBidForOpponent(List<BidDetails> bids) {
                BidDetails maxBid = null;

                for (BidDetails bid : bids) {
                        if (maxBid == null) {
                                maxBid = bid;
                        }
                        else if (this.opponentModel.getBidEvaluation(bid.getBid()) > this.opponentModel.getBidEvaluation(maxBid.getBid())) {
                                maxBid = bid;
                        }
                }

                return maxBid;
        }

        /**
         * Gets called by Negotiator when a relevant negotiation event occurs
         * @param reward
         * @param newState
         */
        public void observeEnvironment(double reward, AbstractState newState) {

                // Only start updating after an action is performed
                // Only update if training is enabled
                if (this.actions.size() > 0 && this.isTraining()) {
                        this.updateQFuction(this.state, this.getLastAction(), reward, newState);
                }
                this.state = newState;
        }

        public HashMap<Integer, ArrayList<Double>> getQTable() {
                return this.qTable;
        }

        protected void updateQFuction(AbstractState state, int action, double reward, AbstractState newState) {
                // initialize state if it is new

                // If agent hasn't done a opening bid, initialize action values to number of bins, otherwise
                // just 3 values (up/down/nothing).
                ArrayList<Double> stateDefaultActionValues = new ArrayList<Double>(Collections.nCopies(state.getActionSize(), 0.0));
                ArrayList<Double> newStateDefaultActionValues = new ArrayList<Double>(Collections.nCopies(newState.getActionSize(), 0.0));

                // Make entries in qTable if they don't exist yet
                this.qTable.putIfAbsent(state.hash(), stateDefaultActionValues);
                this.qTable.putIfAbsent(newState.hash(), newStateDefaultActionValues);


                // To remind ourselves that the below function is correct =>
                // the gamma comes from the domain/preference profile through reward  which is discounted.
                Double gamma = 1.0;
                // Perform update
                Double Qnext = this.maxActionValue(newState);
                Double newActionValue = this.qFunction(state, action) + this.alpha * (reward + gamma * Qnext - this.qFunction(state, action));
                this.qTable.get(state.hash()).set(action, newActionValue);
        }

        /**
         * Determine max_a Q(s,a)
         * @param state The hash of the state for which to retrieve the max action value
         * @return Value of optimal action given that can be taken in the provided state (0 if state is unknown)
         */
        protected Double maxActionValue(AbstractState state) {
                return Collections.max(this.qTable.get(state.hash()));
        }

        /**
         * Get the Q value associated with the provided (state, action) pair.
         * @param state
         * @param action
         * @return
         */
        protected Double qFunction(AbstractState state, int action) {
                ArrayList<Double> actionValues = this.qTable.get(state.hash());
                return actionValues.get(action);
        }

        public void setHyperparameters(StrategyParameters properties) {
                this.eps = properties.getValueAsDouble("epsilon");
                this.alpha = properties.getValueAsDouble("alpha");
                this.bins = (int) properties.getValueAsDouble("bins");
                this.mode = properties.getValueAsString("_mode");
                this.timeBins = (int) properties.getValueAsDouble("time_bins");
        }
        
        protected boolean isTraining() {
                return this.mode.equals("train");
        }
}