示例#1
0
文件: algos.py 项目: manspchr/MADI
def iteration_de_la_valeur(g, p, gamma, epsilon):
    global liste_actions
    start = tm.time()

    #2 etats :  - pour t-1
    #           - pour t
    val_etats = np.zeros((len(g), len(g[0])))
    vt = np.zeros((len(g), len(g[0])))
    q = np.zeros((len(g), len(g[0]), len(liste_actions)))
    t = 0
    #critere d'arret
    while np.max(np.abs(vt - val_etats)) >= epsilon or t == 0:
        t += 1
        #etat precedant
        vt = np.copy(val_etats)
        for i in range(
                len(val_etats)):  #TODO : plus propre de mettre nbLignes non ?
            for j in range(len(val_etats[0])):  # pareil
                for action in liste_actions:
                    r = -g[i][j][1]
                    q[i][j][action] = r + gamma * ot.sum_p_v(
                        g, val_etats, i, j, p, action)
                val_etats[i][j] = max(q[i][j])

    #politique optimal
    d = np.zeros((len(g), len(g[0])))
    for i in range(len(val_etats)):
        for j in range(len(val_etats[0])):
            d[i][j] = np.argmax(q[i][j])
    end = tm.time()
    time = end - start
    return d, val_etats, t, time
示例#2
0
文件: algos.py 项目: manspchr/MADI
def iteration_de_la_politique(g, p, gamma, epsilon):
    global liste_actions

    #2 etats :  - pour t-1
    #           - pour t
    start = tm.time()

    val_etats = np.zeros((len(g), len(g[0])))
    d = np.zeros((len(g), len(g[0])))

    t = 0
    vt = val_etats, val_etats

    #critere d'arret
    while np.max(np.abs(vt - val_etats)) >= epsilon or t == 0:
        #     while not(np.array_equal(d_pred,d)) or t == 0:
        t += 1
        #valeur de la politique courant
        vt = np.copy(val_etats)

        #evaluation de la politique courante
        for i in range(len(val_etats)):
            for j in range(len(val_etats[0])):
                r = -g[i][j][1]
                #arrivé au but
                val_etats[i][j] = r + gamma * ot.sum_p_v(
                    g, val_etats, i, j, p, d[i][j])

        #amelioration de la politique
        for i in range(len(val_etats)):
            for j in range(len(val_etats[0])):
                arg = []
                for action in liste_actions:
                    #arrivé au but
                    r = -g[i][j][1]
                    arg.append(r + gamma *
                               ot.sum_p_v(g, val_etats, i, j, p, action))
                arg = np.array(arg)
                d[i][j] = np.argmax(arg)
#    print(g)
#   print(ot.from_action_to_dir(d,g))
#    #print(v)
#    print(t)
    end = tm.time()
    time = end - start
    return d, val_etats, t, time
示例#3
0
文件: algos.py 项目: danoyu/MADI
def pl(g,p,gamma):
   global liste_actions
   start = tm.time()

   # Create a new model
   m = grb.Model("test")

   val_etats = []

   # Create variables
   for i in range(len(g)):
       for j in range(len(g[0])):
           n = "v["+str(i)+"]["+str(j)+"]"
           r = - g[i][j][1]
           x = m.addVar(vtype=grb.GRB.CONTINUOUS, name=n)
           val_etats.append(x)

   val_etats = np.array(val_etats)
   val_etats = np.reshape(val_etats,(len(g),len(g[0])))

   m.setObjective(np.sum(val_etats), grb.GRB.MINIMIZE)

   for i in range(len(g)):
       for j in range(len(g[0])):
           for action in liste_actions:
               r = -g[i][j][1]
               m.addConstr(val_etats[i][j] >= r + gamma * ot.sum_p_v(g, val_etats, i, j, p, action))

   m.optimize()

   val_etats = np.reshape(np.array([v.x for v in m.getVars()]),(len(g),len(g[0])))

   d = np.zeros((len(g), len(g[0])))
   for i in range(len(val_etats)):
       for j in range(len(val_etats[0])):
           val_par_action = np.zeros(len(liste_actions))
           for action in liste_actions:
                   val_par_action[action] = r + gamma * ot.sum_p_v(g, val_etats, i, j, p, action)
           d[i][j] = np.argmax(val_par_action)
   end = tm.time()
   time = end - start
   #nombre d'iteration
   t = m.getAttr('IterCount')
   return d, val_etats, t, time
示例#4
0
文件: test.py 项目: manspchr/MADI
def pl(g, nbLignes, nbCol, liste_actions, rewards, p, gamma, epsilon):

    # Create a new model
    m = Model("test")

    val_etats = []

    # Create variables
    for i in range(nbLignes):
        for j in range(nbCol):
            n = "v[" + str(i) + "][" + str(j) + "]"
            r = rewards[g[i][j]]
            if i == len(g) - 1 and j == len(g[0]) - 1:
                r = rewards[-1]
            x = m.addVar(vtype=GRB.CONTINUOUS, name=n)
            val_etats.append(x)

    # Set objective
    # print(val_etats[0])
    # print(m.getVars())
    # for v in m.getVars():
    #     print('la')
    #     print('%s %g' % (v.varName, v.x))
    # print(y)

    val_etats = np.array(val_etats)
    val_etats = np.reshape(val_etats, (nbLignes, nbCol))

    m.setObjective(np.sum(val_etats), GRB.MINIMIZE)
    print((val_etats.shape))
    for i in range(nbLignes):
        for j in range(nbCol):
            for action in liste_actions:
                r = outils.reward_action(g, i, j, action, rewards)
                if r:
                    if i == len(g) - 1 and j == len(g[0]) - 1:
                        r = rewards[-1]
                    m.addConstr(
                        val_etats[i][j] >= r +
                        gamma * outils.sum_p_v(g, val_etats, i, j, p, action))

    m.optimize()

    # for v in m.getVars():
    #     print('%s %g' % (v.varName, v.x))
    val_etats = np.reshape(np.array([v.x for v in m.getVars()]),
                           (nbLignes, nbCol))
    # print(val_etats)
    # print(val_etats.shape)

    # for v in m.getConstrs():
    #     print('%s' % (v.getAttr("slack")))

    # print('Obj: %g' % m.objVal)

    # Affiche le pl dans un fichier
    # m.write('debug.lp')

    d = np.zeros((nbLignes, nbCol))
    for i in range(len(val_etats)):
        for j in range(len(val_etats[0])):
            val_par_action = np.zeros(len(liste_actions))
            for action in liste_actions:
                val_par_action[action] = r + gamma * outils.sum_p_v(
                    g, val_etats, i, j, p, action)
            d[i][j] = np.argmax(val_par_action)
    return d, val_etats