Jubatus0.7.0で実装された多腕バンディットを使う

使う関数

　register_arm([腕（string）])
　register_reward([腕を回す人（string）], [腕（string）], [報酬（float）])
　[次に回す腕(string)] = select_arm([腕を回す人（string）])

Config(UCB1の場合)

ucb1 softmax exp3 epsilon_greedyがあるっぽい

ucb1.json

 {
  "method": "ucb1",
  "parameter": {
  }
}

起動

jubatus-0.7.0/config/bandit/の中にサンプルjsonがあるので使うのが楽

jubabandit -f /usr/local/src/jubatus-installer-master/download/jubatus-0.7.0/config/bandit/ucb1.json

pythonサンプル

bandit.py

#!/usr/bin/env python
# coding: utf-8

host = '127.0.0.1'
port = 9199
name = 'test'

import sys
import json
import random

import jubatus
from jubatus.common import Datum

def train(client):
    arms = [
        u'ヤフー',
        u'グーグル',
        u'楽天',
    ]
    player = u'bandit'

    for arm in arms:
        client.register_arm(arm)
        reword = 0.0 if random.randint(1, 100) <= 70 else 1.0
        client.register_reward(player, arm, reword)
        
        sys.stdout.write(arm.encode('utf-8'))
        sys.stdout.write('\n')
        sys.stdout.write(str(reword))
        sys.stdout.write('\n')

def predict(client):
    player = u'bandit'
    res = client.select_arm(player)
    sys.stdout.write(str(res))
    sys.stdout.write('\n')

if __name__ == '__main__':
    # connect to the jubatus
    client = jubatus.Bandit(host, port, name)
    # run example
    train(client)
    predict(client)