/* Copyright 2009 by Marcin Szubert Licensed under the Academic Free License version 3.0 */ package cecj.eval; import cecj.interaction.InteractionResult; import cecj.problems.SymmetricCompetitionProblem; import ec.EvolutionState; import ec.Individual; import ec.simple.SimpleFitness; import ec.util.Parameter; /** * Single elimination tournament competitive evaluator. It is different from the other * coevolutionary evaluators because interactions between individuals must be simulated in strict * order. It depends on the outcome of previous interaction if the individual can compete further. * * Assumes that individuals use SimpleFitness. The fitness assigned by this method is * equal to the height of the tournament subtree that particular individual has traversed - the * number of games won. To reduce the inherent noise of the tournament evaluation scheme, a few * rounds can be played. The number of rounds is specified by a repeats parameter which * is equal to 1 by default. This evaluator can be used if problem being solved implements * SymmetricCompetitionProblem interface. * * Since it would be hard to extend this evaluator with generic archiving or fitness sharing, only * the simplest settings are available. * * @author Marcin Szubert * */ public class TournamentCoevolutionaryEvaluator extends CoevolutionaryEvaluator { private static final String P_REPEATS = "repeats"; /** * Specifies how many times the tournament should be repeated during single evaluation process. * More repeats can reduce the noise of this evaluation scheme. */ private int tournamentRepeats; private SymmetricCompetitionProblem problem; /** * Represents competing individuals. */ private Individual[] competitors; /** * Number of competitors - size of the particular subpopulation. */ private int numCompetitors; /** * Points gathered during the course of competition. */ private int[] points; /** * An array used as a tournament tree representation. It stores indices of competing * individuals. Neighboring indices compete with each other in certain round. */ private int[] competition; /** * Stores active competitors ready to be divided into pairs. */ private int[] activeCompetitors; /** * Indicates if particular competitor is still in game. */ private boolean[] active; @Override public void setup(final EvolutionState state, final Parameter base) { super.setup(state, base); if (!(p_problem instanceof SymmetricCompetitionProblem)) { state.output.fatal("Tournament evaluator can be used only with symmetric problems"); } else { problem = (SymmetricCompetitionProblem) p_problem; } Parameter repeatsParameter = base.push(P_REPEATS); tournamentRepeats = state.parameters.getIntWithDefault(repeatsParameter, null, 1); if (tournamentRepeats <= 0) { state.output.fatal("Tournament repeats parameter can not be negative.", repeatsParameter); } } @Override public void evaluatePopulation(EvolutionState state) { for (int subpop = 0; subpop < numSubpopulations; subpop++) { prepareTournament(state, subpop); for (int r = 0; r < tournamentRepeats; r++) { makeTournament(state); } assignFitness(state); } } /** * Initializes structures used in the tournament series. * * @param state * current evolutionary state * @param subpop * index of subpopulation */ private void prepareTournament(EvolutionState state, int subpop) { competitors = state.population.subpops[subpop].individuals; numCompetitors = competitors.length; points = new int[numCompetitors]; active = new boolean[numCompetitors]; competition = new int[numCompetitors]; activeCompetitors = new int[numCompetitors]; } /** * Plays a single tournament between earlier selected competitors from particular subpopulation. * Each tournament consists of a sequence of rounds. In each round number of active competitors * is reduced by half according to the results of their games (approximately - if at the start * of the round number of players is odd, one player is given a "bye" and advances to the next * round directly). At the beginning of each round there is a drawing which assigns competitors * in pairs. * * @param state * current evolutionary state */ private void makeTournament(EvolutionState state) { int numActiveCompetitors; for (int c = 0; c < numCompetitors; c++) { active[c] = true; } while ((numActiveCompetitors = findActiveCompetitors()) > 1) { shuffleCompetitors(state, numActiveCompetitors); playTournamentRound(state, numActiveCompetitors); } } /** * Assigns fitness value to each competing individual according to overall points which it has * gathered during the series of tournaments. * * @param state * current evolutionary state */ private void assignFitness(EvolutionState state) { for (int c = 0; c < numCompetitors; c++) { Individual competitor = competitors[c]; ((SimpleFitness) competitor.fitness).setFitness(state, points[c], false); } } /** * Finds still active competitors according to active array. Found competitor * indices are stored in activeCompetitors array and their number is returned. * * @return number of still active competitors */ private int findActiveCompetitors() { int leftCompetitors = 0; for (int c = 0; c < numCompetitors; c++) { if (active[c]) { activeCompetitors[leftCompetitors++] = c; } } return leftCompetitors; } /** * Randomly shuffles competitors indices taken from < * * @param state * current evolutionary state * @param count * the number of shuffled competitors */ private void shuffleCompetitors(EvolutionState state, int count) { int left = count; for (int i = 0; i < count; i++) { int rand = state.random[0].nextInt(left); competition[i] = activeCompetitors[rand]; activeCompetitors[rand] = activeCompetitors[--left]; } } /** * Arranges a competition between neighbors in competition array. * * @param state * current evolutionary state * @param numLeftCompetitors * the number of competitors left */ private void playTournamentRound(EvolutionState state, int numLeftCompetitors) { for (int i = 0; i + 1 < numLeftCompetitors; i += 2) { Individual c1 = competitors[competition[i]]; Individual c2 = competitors[competition[i + 1]]; // TODO: consider if it is needed to call compete method twice // maybe it should use internal individual's fitness or return both // results at once? InteractionResult score1 = problem.compete(state, c1, c2).first; InteractionResult score2 = problem.compete(state, c2, c1).first; if (score1.betterThan(score2)) { points[competition[i]]++; active[competition[i + 1]] = false; } else { points[competition[i + 1]]++; active[competition[i]] = false; } } // TODO: in case of odd number of competitors, should the one given a // "bye" achieve a point // in this round? if (numLeftCompetitors % 2 != 0) { points[competition[numLeftCompetitors - 1]]++; } } }