Value Iteration Frozen Lake.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Reinforcement Learning"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import gym\n",
    "import collections\n",
    "from tensorboardX import SummaryWriter"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "ENV_NAME = \"FrozenLake-v0\"\n",
    "GAMMA = 0.9\n",
    "TEST_EPISODES = 20"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Agent:\n",
    "    def __init__(self):\n",
    "        self.env = gym.make(ENV_NAME) # create Frozen Lake\n",
    "        self.state = self.env.reset() # Starting state.\n",
    "        self.rewards = collections.defaultdict(float) # Reward Matrix\n",
    "        self.transits = collections.defaultdict(collections.Counter) # trasition probs.\n",
    "        self.values = collections.defaultdict(float) # state values. Vs\n",
    "\n",
    "    def play_n_random_steps(self, count):\n",
    "        for _ in range(count):\n",
    "            action = self.env.action_space.sample() # Get a random action\n",
    "            new_state, reward, is_done, _ = self.env.step(action) # Perform the action\n",
    "            self.rewards[(self.state, action, new_state)] = reward # make an entry in reward table.\n",
    "            self.transits[(self.state, action)][new_state] += 1 # transition count, later converted to probability.\n",
    "            self.state = self.env.reset() if is_done else new_state # if episode done reset the environment else new state.\n",
    "\n",
    "    def calc_action_value(self, state, action):\n",
    "        target_counts = self.transits[(state, action)] # Get transition counts.\n",
    "        total = sum(target_counts.values()) # Total transitions from state\n",
    "        action_value = 0.0\n",
    "        for tgt_state, count in target_counts.items(): # for each target state on doing action a instate s.\n",
    "            reward = self.rewards[(state, action, tgt_state)] # Get reward.\n",
    "            action_value += (count / total) * (reward + GAMMA * self.values[tgt_state]) # Action value for a state-action pair.\n",
    "        return action_value\n",
    "\n",
    "    def select_action(self, state):\n",
    "        best_action, best_value = None, None\n",
    "        for action in range(self.env.action_space.n): # for each action.\n",
    "            action_value = self.calc_action_value(state, action) # For each state-action pair get action value Vs\n",
    "            if best_value is None or best_value < action_value: # decide best action.\n",
    "                best_value = action_value\n",
    "                best_action = action\n",
    "        return best_action\n",
    "\n",
    "    def play_episode(self, env):\n",
    "        total_reward = 0.0\n",
    "        state = env.reset()\n",
    "        while True:\n",
    "            action = self.select_action(state) # Choose Best Action.\n",
    "            new_state, reward, is_done, _ = env.step(action) # Do Action.\n",
    "            self.rewards[(state, action, new_state)] = reward # Get reward.\n",
    "            self.transits[(state, action)][new_state] += 1 # Make transition entry +1.\n",
    "            total_reward += reward\n",
    "            if is_done:\n",
    "                break\n",
    "            state = new_state\n",
    "        return total_reward\n",
    "\n",
    "    def value_iteration(self):\n",
    "        for state in range(self.env.observation_space.n): # For every state.\n",
    "            state_values = [self.calc_action_value(state, action) # Get state-action values.\n",
    "                            for action in range(self.env.action_space.n)]\n",
    "            self.values[state] = max(state_values) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Best reward updated 0.000 -> 0.050\n",
      "Best reward updated 0.050 -> 0.100\n",
      "Best reward updated 0.100 -> 0.300\n",
      "Best reward updated 0.300 -> 0.350\n",
      "Best reward updated 0.350 -> 0.500\n",
      "Best reward updated 0.500 -> 0.550\n",
      "Best reward updated 0.550 -> 0.650\n",
      "Best reward updated 0.650 -> 0.750\n",
      "Best reward updated 0.750 -> 0.800\n",
      "Best reward updated 0.800 -> 0.850\n",
      "Solved in 88 iterations!\n"
     ]
    }
   ],
   "source": [
    "if __name__ == \"__main__\":\n",
    "    test_env = gym.make(ENV_NAME)\n",
    "    agent = Agent()\n",
    "    writer = SummaryWriter(comment=\"-v-iteration\")\n",
    "\n",
    "    iter_no = 0\n",
    "    best_reward = 0.0\n",
    "    while True:\n",
    "        iter_no += 1\n",
    "        agent.play_n_random_steps(100)\n",
    "        agent.value_iteration()\n",
    "\n",
    "        reward = 0.0\n",
    "        for _ in range(TEST_EPISODES):\n",
    "            reward += agent.play_episode(test_env)\n",
    "        reward /= TEST_EPISODES\n",
    "        writer.add_scalar(\"reward\", reward, iter_no)\n",
    "        if reward > best_reward:\n",
    "            print(\"Best reward updated %.3f -> %.3f\" % (best_reward, reward))\n",
    "            best_reward = reward\n",
    "        if reward > 0.80:\n",
    "            print(\"Solved in %d iterations!\" % iter_no)\n",
    "            break\n",
    "    writer.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}