• +91-9872993883
• +91-8283824812
• info@ris-ai.com

# Mountain Car Problem¶

## What is mountain Car Problem? ¶

A car is on a one-dimensional track, positioned between two mountains. The goal is to drive up the mountain on the right (reaching the flag). However, the car’s engine is not strong enough to climb the mountain in a single pass. Therefore, the only way to succeed is to drive back and forth to build up momentum.

### How mountain car works? ¶

A car is an agent. The environment is a black-box world of one-dimensional mountains. The car’s action boils down to only one number: if positive, the engine pushes the car to the right. If negative, it pushes the car to the left. The agent perceives an environment through an observation: the car’s X position and velocity. If we want our car to drive on top of the mountain, we define the reward in a convenient way: The agent gets -1 to its reward for every step in which it hasn’t reached the goal. When it reaches the goal, the episode ends. So, in fact, the agent is punished for not being in a position we want it to be. The faster he reaches it, the better for him. The agent’s goal is to maximize the total reward, which is the sum of rewards from one episode. So if it reaches the desired point after, e.g., 110 steps, it receives a total return of -110, which would be a great result for Mountain Car, because if it doesn’t reach the goal, then it is punished for 200 steps (hence, a return of -200).

This is the whole problem formulation. Now, we can give it to the algorithms, which are already powerful enough to solve such problems in a matter of minute. It’s worth noting that we don’t tell the agent how to achieve the goal. We don’t even provide any hints (heuristics). The agent will find a way (a policy) to win on its own.

#### 1. Importing Different Libraries ¶

We don’t need to implement the Mountain Car environment ourselves; the OpenAI Gym library provides that implementation. Let’s see a random agent (an agent that takes random actions) in our environment:

In [ ]:
import numpy as np
import gym
from keras.models import Sequential
from keras.layers import Dense, Activation, Flatten
from rl.agents.dqn import DQNAgent
from rl.policy import BoltzmannQPolicy, EpsGreedyQPolicy
from rl.memory import SequentialMemory
ENV_NAME = 'MountainCar-v0'
# Get the environment and extract the number of actions.
env = gym.make(ENV_NAME)

In [ ]:
## play for rondom action, without being trained
for i_episode in range(5):
observation = env.reset()
for t in range(100):
env.render()
print(observation)
action = env.action_space.sample()
observation, reward, done, info = env.step(action)
if done:
print("Episode finished after {} timesteps".format(t+1))
break
env.close()

In [ ]:
# Get the environment and extract the number of actions.
env = gym.make(ENV_NAME)
np.random.seed(123)
env.seed(123)
nb_actions = env.action_space.n
# Next, we build a very simple model.
model = Sequential()

#
print(model.summary())

In [ ]:
# # Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# # even the metrics!
memory = SequentialMemory(limit=50000, window_length=1)
policy = EpsGreedyQPolicy()
dqn = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, nb_steps_warmup=50,
target_model_update=200,train_interval=4, policy=policy)
#

In [ ]:
# # Okay, now it's time to learn something! We visualize the training here for show, but this
# # slows down training quite a lot. You can always safely abort the training prematurely using
# # Ctrl + C.
### uncomment this section to train your model,
# dqn.fit(env, nb_steps=10000, visualize=False, verbose=2)
#
# # Uncomment this to save your own weight
# dqn.save_weights('dqn_{}_weights.h5f'.format(ENV_NAME), overwrite=True)

In [ ]:
#While training comment below two line
#
weights_filename = 'dqn_{}_weights.h5f'.format(ENV_NAME)

dqn.test(env, nb_episodes=10, visualize=True)

env.close()