IIT-M RL-ASSIGNMENT-2-TAXI
Solution for submission 132232
A detailed solution for submission 132232 submitted for challenge IIT-M RL-ASSIGNMENT-2-TAXI
What is the notebook about?¶
Problem - Taxi Environment Algorithms¶
This problem deals with a taxi environment and stochastic actions. The tasks you have to do are:
- Implement Policy Iteration
- Implement Modified Policy Iteration
- Implement Value Iteration
- Implement Gauss Seidel Value Iteration
- Visualize the results
- Explain the results
How to use this notebook? 📝¶
This is a shared template and any edits you make here will not be saved.You should make a copy in your own drive. Click the "File" menu (top-left), then "Save a Copy in Drive". You will be working in your copy however you like.
Update the config parameters. You can define the common variables here
Variable | Description |
---|---|
AICROWD_DATASET_PATH |
Path to the file containing test data. This should be an absolute path. |
AICROWD_RESULTS_DIR |
Path to write the output to. |
AICROWD_ASSETS_DIR |
In case your notebook needs additional files (like model weights, etc.,), you can add them to a directory and specify the path to the directory here (please specify relative path). The contents of this directory will be sent to AIcrowd for evaluation. |
AICROWD_API_KEY |
In order to submit your code to AIcrowd, you need to provide your account's API key. This key is available at https://www.aicrowd.com/participants/me |
- Installing packages. Please use the Install packages 🗃 section to install the packages
Setup AIcrowd Utilities 🛠¶
We use this to bundle the files for submission and create a submission on AIcrowd. Do not edit this block.
!pip install aicrowd-cli > /dev/null
AIcrowd Runtime Configuration 🧷¶
Get login API key from https://www.aicrowd.com/participants/me
import os
from matplotlib import pyplot as plt
AICROWD_DATASET_PATH = os.getenv("DATASET_PATH", os.getcwd()+"/13d77bb0-b325-4e95-a03b-833eb6694acd_a2_taxi_inputs.zip")
AICROWD_RESULTS_DIR = os.getenv("OUTPUTS_DIR", "results")
!unzip $AICROWD_DATASET_PATH
DATASET_DIR = 'inputs/'
Taxi Environment¶
Read the environment to understand the functions, but do not edit anything
import numpy as np
class TaxiEnv_HW2:
def __init__(self, states, actions, probabilities, rewards, initial_policy):
self.possible_states = states
self._possible_actions = {st: ac for st, ac in zip(states, actions)}
self._ride_probabilities = {st: pr for st, pr in zip(states, probabilities)}
self._ride_rewards = {st: rw for st, rw in zip(states, rewards)}
self.initial_policy = initial_policy
self._verify()
def _check_state(self, state):
assert state in self.possible_states, "State %s is not a valid state" % state
def _verify(self):
"""
Verify that data conditions are met:
Number of actions matches shape of next state and actions
Every probability distribution adds up to 1
"""
ns = len(self.possible_states)
for state in self.possible_states:
ac = self._possible_actions[state]
na = len(ac)
rp = self._ride_probabilities[state]
assert np.all(rp.shape == (na, ns)), "Probabilities shape mismatch"
rr = self._ride_rewards[state]
assert np.all(rr.shape == (na, ns)), "Rewards shape mismatch"
assert np.allclose(rp.sum(axis=1), 1), "Probabilities don't add up to 1"
def possible_actions(self, state):
""" Return all possible actions from a given state """
self._check_state(state)
return self._possible_actions[state]
def ride_probabilities(self, state, action):
"""
Returns all possible ride probabilities from a state for a given action
For every action a list with the returned with values in the same order as self.possible_states
"""
actions = self.possible_actions(state)
ac_idx = actions.index(action)
return self._ride_probabilities[state][ac_idx]
def ride_rewards(self, state, action):
actions = self.possible_actions(state)
ac_idx = actions.index(action)
return self._ride_rewards[state][ac_idx]
Example of Environment usage¶
def check_taxienv():
# These are the values as used in the pdf, but they may be changed during submission, so do not hardcode anything
states = ['A', 'B', 'C']
actions = [['1','2','3'], ['1','2'], ['1','2','3']]
probs = [np.array([[1/2, 1/4, 1/4],
[1/16, 3/4, 3/16],
[1/4, 1/8, 5/8]]),
np.array([[1/2, 0, 1/2],
[1/16, 7/8, 1/16]]),
np.array([[1/4, 1/4, 1/2],
[1/8, 3/4, 1/8],
[3/4, 1/16, 3/16]]),]
rewards = [np.array([[10, 4, 8],
[ 8, 2, 4],
[ 4, 6, 4]]),
np.array([[14, 0, 18],
[ 8, 16, 8]]),
np.array([[10, 2, 8],
[6, 4, 2],
[4, 0, 8]]),]
initial_policy = {'A': '1', 'B': '1', 'C': '1'}
env = TaxiEnv_HW2(states, actions, probs, rewards, initial_policy)
print("All possible states", env.possible_states)
print("All possible actions from state B", env.possible_actions('B'))
print("Ride probabilities from state A with action 2", env.ride_probabilities('A', '2'))
print("Ride rewards from state C with action 3", env.ride_rewards('C', '3'))
base_kwargs = {"states": states, "actions": actions,
"probabilities": probs, "rewards": rewards,
"initial_policy": initial_policy}
return base_kwargs
base_kwargs = check_taxienv()
env = TaxiEnv_HW2(**base_kwargs)
Task 1 - Policy Iteration¶
Run policy iteration on the environment and generate the policy and expected reward
# 1.1 Policy Iteration
def policy_iteration(taxienv, gamma):
# A list of all the states
states = taxienv.possible_states
# Initial values
values = {s: 0 for s in states}
# This is a dictionary of states to policies -> e.g {'A': '1', 'B': '2', 'C': '1'}
policy = taxienv.initial_policy.copy()
## Begin code here
# Hints -
# Do not hardcode anything
# Only the final result is required for the results
# Put any extra data in "extra_info" dictonary for any plots etc
# Use the helper functions taxienv.ride_rewards, taxienv.ride_probabilities, taxienv.possible_actions
# For terminating condition use the condition exactly mentioned in the pdf
actions=taxienv._possible_actions
ride_probs=taxienv._ride_probabilities
ride_rewds=taxienv._ride_rewards
done,time_local,time_global,unequal_policy=0,0,0,0
extra_info = {}
extra_info['time_local'],extra_info['time_global'],extra_info['delta'],extra_info['unequal_policy']=[],[],[],[]
while(done==0):
time_global+=1
delta=1e-8
while(delta>=1e-8):
time_local+=1
delta=0
for state in states:
j=values[state]
action=policy[state]
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
values[state]=sum(ride_prob*(ride_rewd+gamma*values_)) #sum(taxienv.ride_probabilities(state,action)*(env.ride_rewards(state,action)+gamma*np.array(list(values.values()))))
delta=max(delta,abs(j-values[state]))
extra_info['delta'].append(delta)
extra_info['time_local'].append(time_local)
time_local=0
done=1
for state in states:
b=policy[state]
t_j=[]
for action in actions[state]:
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
t_j.append(sum(ride_prob*(ride_rewd+gamma*values_)))
policy[state]=actions[state][np.argmax(np.array(t_j))]
if b!=policy[state]:
done=0
if len(extra_info['unequal_policy'])<time_global:
extra_info['unequal_policy'].append(0)
extra_info['unequal_policy'][time_global-1]+=1
extra_info['time_global'].append(time_global)
## Do not edit below this line
# Final results
return {"Expected Reward": values, "Policy": policy}, extra_info
Task 2 - Policy Iteration for multiple values of gamma¶
Ideally this code should run as is
def run_policy_iteration(env):
gamma_values = np.arange(5, 100, 5)/100
results, extra_info = {}, {}
for gamma in gamma_values:
results[gamma], extra_info[gamma] = policy_iteration(env, gamma)
return results, extra_info
results, extra_info = run_policy_iteration(env)
Task 3 - Modifed Policy Iteration¶
Implement modified policy iteration (where Value iteration is done for fixed m number of steps)
# 1.3 Modified Policy Iteration
def modified_policy_iteration(taxienv, gamma, m):
# A list of all the states
states = taxienv.possible_states
# Initial values
values = {s: 0 for s in states}
# This is a dictionary of states to policies -> e.g {'A': '1', 'B': '2', 'C': '1'}
policy = taxienv.initial_policy.copy()
## Begin code here
# Hints -
# Do not hardcode anything
# Only the final result is required for the results
# Put any extra data in "extra_info" dictonary for any plots etc
# Use the helper functions taxienv.ride_rewards, taxienv.ride_probabilities, taxienv.possible_actions
# For terminating condition use the condition exactly mentioned in the pdf
actions=taxienv._possible_actions
ride_probs=taxienv._ride_probabilities
ride_rewds=taxienv._ride_rewards
done,time_local,time_global,unequal_policy=0,0,0,0
extra_info = {}
extra_info['time_local'],extra_info['time_global'],extra_info['unequal_policy']=[],[],[]
while(done==0):
time_global+=1
for z in range(m+1):
time_local+=1
h={}
for state in states:
action=policy[state]
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
h[state]=sum(ride_prob*(ride_rewd+gamma*values_))
values=h.copy()
extra_info['time_local'].append(time_local)
time_local=0
done=1
for state in states:
b=policy[state]
t_j=[]
for action in actions[state]:
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
t_j.append(sum(ride_prob*(ride_rewd+gamma*values_)))
policy[state]=actions[state][np.argmax(np.array(t_j))]
if b!=policy[state]:
done=0
if len(extra_info['unequal_policy'])<time_global:
extra_info['unequal_policy'].append(0)
extra_info['unequal_policy'][time_global-1]+=1
extra_info['time_global'].append(time_global)
## Do not edit below this line
# Final results
return {"Expected Reward": values, "Policy": policy}, extra_info
Task 4 Modified policy iteration for multiple values of m¶
Ideally this code should run as is
def run_modified_policy_iteration(env):
m_values = np.arange(1, 15)
gamma = 0.9
results, extra_info = {}, {}
for m in m_values:
results[m], extra_info[m] = modified_policy_iteration(env, gamma, m)
return results, extra_info
results, extra_info = run_modified_policy_iteration(env)
#print('modified PI results, extra_info: ',results,extra_info)
Task 5 Value Iteration¶
Implement value iteration and find the policy and expected rewards
# 1.4 Value Iteration
def value_iteration(taxienv, gamma):
# A list of all the states
states = taxienv.possible_states
# Initial values
values = {s: 0 for s in states}
# This is a dictionary of states to policies -> e.g {'A': '1', 'B': '2', 'C': '1'}
policy = taxienv.initial_policy.copy()
## Begin code here
# Hints -
# Do not hardcode anything
# Only the final result is required for the results
# Put any extra data in "extra_info" dictonary for any plots etc
# Use the helper functions taxienv.ride_rewards, taxienv.ride_probabilities, taxienv.possible_actions
# For terminating condition use the condition exactly mentioned in the pdf
actions=taxienv._possible_actions
ride_probs=taxienv._ride_probabilities
ride_rewds=taxienv._ride_rewards
delta,time_global=1e-8,0
extra_info = {}
extra_info['time_global'],extra_info['delta']=[],[]
while(delta>=1e-8):
time_global+=1
delta=0
h={}
for state in states:
t_j=[]
for action in actions[state]:
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
t_j.append(sum(ride_prob*(ride_rewd+gamma*values_)))
h[state]=max(t_j)
policy[state]=actions[state][np.argmax(np.array(t_j))]
delta=max(delta,abs(values[state]-h[state]))
extra_info['delta'].append(delta)
values=h.copy()
extra_info['time_global'].append(time_global)
## Do not edit below this line
# Final results
return {"Expected Reward": values, "Policy": policy}, extra_info
Task 6 Value Iteration with multiple values of gamma¶
Ideally this code should run as is
def run_value_iteration(env):
gamma_values = np.arange(5, 100, 5)/100
results = {}
results, extra_info = {}, {}
for gamma in gamma_values:
results[gamma], extra_info[gamma] = value_iteration(env, gamma)
return results, extra_info
results, extra_info = run_value_iteration(env)
Task 7 Gauss Seidel Value Iteration¶
Implement Gauss Seidel Value Iteration
# 1.4 Gauss Seidel Value Iteration
def gauss_seidel_value_iteration(taxienv, gamma):
# A list of all the states
# For Gauss Seidel Value Iteration - iterate through the values in the same order
states = taxienv.possible_states
# Initial values
values = {s: 0 for s in states}
# This is a dictionary of states to policies -> e.g {'A': '1', 'B': '2', 'C': '1'}
policy = taxienv.initial_policy.copy()
# Hints -
# Do not hardcode anything
# For Gauss Seidel Value Iteration - iterate through the values in the same order as taxienv.possible_states
# Only the final result is required for the results
# Put any extra data in "extra_info" dictonary for any plots etc
# Use the helper functions taxienv.ride_rewards, taxienv.ride_probabilities, taxienv.possible_actions
# For terminating condition use the condition exactly mentioned in the pdf
## Begin code here
actions=taxienv._possible_actions
ride_probs=taxienv._ride_probabilities
ride_rewds=taxienv._ride_rewards
delta,time_global=1e-8,0
extra_info = {}
extra_info['time_global'],extra_info['delta']=[],[]
while(delta>=1e-8):
time_global+=1
delta=0
h={}
for state in states:
j=values[state]
t_j=[]
for action in actions[state]:
index=actions[state].index(action)
ride_prob=ride_probs[state][index]
ride_rewd=ride_rewds[state][index]
values_=np.array(list(values.values()))
t_j.append(sum(ride_prob*(ride_rewd+gamma*values_)))
values[state]=max(t_j)
policy[state]=actions[state][np.argmax(np.array(t_j))]
delta=max(delta,abs(values[state]-j))
extra_info['delta'].append(delta)
extra_info['time_global'].append(time_global)
## Do not edit below this line
# Final results
return {"Expected Reward": values, "Policy": policy}, extra_info
Task 8 Gauss Seidel Value Iteration with multiple values of gamma¶
Ideally this code should run as is
def run_gauss_seidel_value_iteration(env):
gamma_values = np.arange(5, 100, 5)/100
results = {}
results, extra_info = {}, {}
for gamma in gamma_values:
results[gamma], extra_info[gamma] = gauss_seidel_value_iteration(env, gamma)
return results, extra_info
results, extra_info = run_gauss_seidel_value_iteration(env)
'''
###comparing and contrasting value iterations and gauss seidel value iteration###
gamma_values = np.arange(5, 100, 5)/100
for gamma in gamma_values:
print('vi: ',results1[gamma],extra_info1[gamma]) #1 stands for VI results, extra_info
print('gsvi: ',results2[gamma],extra_info2[gamma]) #2 stands for GSVI results, extra_info
'''
Generate Results ✅¶
# Do not edit this cell
def get_results(kwargs):
taxienv = TaxiEnv_HW2(**kwargs)
policy_iteration_results = run_policy_iteration(taxienv)[0]
modified_policy_iteration_results = run_modified_policy_iteration(taxienv)[0]
value_iteration_results = run_value_iteration(taxienv)[0]
gs_vi_results = run_gauss_seidel_value_iteration(taxienv)[0]
final_results = {}
final_results["policy_iteration"] = policy_iteration_results
final_results["modifed_policy_iteration"] = modified_policy_iteration_results
final_results["value_iteration"] = value_iteration_results
final_results["gauss_seidel_iteration"] = gs_vi_results
return final_results
# Do not edit this cell, generate results with it as is
if not os.path.exists(AICROWD_RESULTS_DIR):
os.mkdir(AICROWD_RESULTS_DIR)
for params_file in os.listdir(DATASET_DIR):
kwargs = np.load(os.path.join(DATASET_DIR, params_file), allow_pickle=True).item()
results = get_results(kwargs)
idx = params_file.split('_')[-1][:-4]
np.save(os.path.join(AICROWD_RESULTS_DIR, 'results_' + idx), results)
Check your local score¶
This score is not your final score, and it doesn't use the marks weightages. This is only for your reference of how arrays are matched and with what tolerance.
# Check your score on the given test cases (There are more private test cases not provided)
target_folder = 'targets'
result_folder = AICROWD_RESULTS_DIR
def check_algo_match(results, targets):
param_matches = []
for k in results:
param_results = results[k]
param_targets = targets[k]
policy_match = param_results['Policy'] == param_targets['Policy']
rv = [v for k, v in param_results['Expected Reward'].items()]
tv = [v for k, v in param_targets['Expected Reward'].items()]
rewards_match = np.allclose(rv, tv, rtol=3)
equal = rewards_match and policy_match
param_matches.append(equal)
return np.mean(param_matches)
def check_score(target_folder, result_folder):
match = []
for out_file in os.listdir(result_folder):
res_file = os.path.join(result_folder, out_file)
results = np.load(res_file, allow_pickle=True).item()
idx = out_file.split('_')[-1][:-4] # Extract the file number
target_file = os.path.join(target_folder, f"targets_{idx}.npy")
targets = np.load(target_file, allow_pickle=True).item()
algo_match = []
for k in targets:
algo_results = results[k]
algo_targets = targets[k]
algo_match.append(check_algo_match(algo_results, algo_targets))
match.append(np.mean(algo_match))
return np.mean(match)
if os.path.exists(target_folder):
print("Shared data Score (normalized to 1):", check_score(target_folder, result_folder))
Visualize results of Policy Iteration with multiple values of gamma¶
Add code to visualize the results
## Visualize policy iteration with multiple values of gamma
#'results', 'extra_info' variables here stand for the outputs from 'policy iteraton' (1.2) and not others (as they also have same variable names)
#So, run the below commented section just after running 'policy iteration' (Task 1) for multiple values of gamma
'''
gamma_values = np.arange(5, 100, 5)/100
local_time,global_time,delta=[],[],[]
values=[]
twos=[]
for gamma in gamma_values:
local_time.append(extra_info[gamma]['time_local'][-1])
global_time.append(extra_info[gamma]['time_global'][-1])
delta.append(extra_info[gamma]['delta'][-1])
values.append(list(results[gamma]["Expected Reward"].values()))
twos.append(list(results[gamma]["Policy"].values()).count('2'))
plt.plot(gamma_values,local_time)
plt.xlabel('gamma values')
plt.ylabel('No. of iterations for inner (delta<1e-8) loop')
plt.title('gamma vs no. of inner loop iterations')
plt.show()
plt.plot(gamma_values,global_time)
plt.xlabel('gamma values')
plt.ylabel('No. of iterations for outer (done) loop')
plt.title('gamma vs no. of outer loop iterations')
plt.show()
for i in range(len(gamma_values)):
plt.plot(['A','B','C'],values[i],label=gamma_values[i])
plt.xlabel('gamma values')
plt.ylabel('Cummulative values for 3 states: A, B, C')
plt.title('gamma vs expected values')
plt.legend(loc="upper right")
plt.show()
plt.plot(delta)
plt.xlabel('gamma*no. of iterations')
plt.ylabel('delta')
plt.title('delta vs multiple iterations of different gamma')
plt.show()
plt.plot(gamma_values,twos)
plt.xlabel('gamma')
plt.ylabel('no. of states (out of 3) with action 2')
plt.title('no. of states with action 2 vs gamma')
plt.show()
'''
Subjective questions¶
1.a How are values of $\gamma$ affecting results of policy iteration¶
As can be observed from the graphs, as gamma increases, number of inner loop iterations (policy improvement step), number of outer loop iterations (policy evaluation and repeating), maximum delta values, expected values for any state ('A','B','C') increases as more and more priority is given to future-transitions/long-term consideration. Also, starting from 0 states which had action 2 as optimal action (when gamma=0) as greedily action 2 is worse as action 1 has more per-stage expected reward from any state, we finally end up with a policy that has all 3 optimal actions (ie. for all 3 states) being 2 as when analysed in long term, action 2 at any state has more probability of ending up at B, which has a good expected reward at B and very high (3/4) transition probability to B again.
1.b For modified policy itetaration, do you find any improvement if you choose m=10.¶
No considerable improvement is observed if m>10 (value difference for any state is atmost 2) as can be observed from results and 'extra-info' obtained from 1.3. This might be because the policy(/value) improvement step is a contraction step and it converges quickly (which occurs for m<10) in our case!
1.c Compare and contrast the behavior of Value Iteration (VI) and Gauss Seidel Value Iteraton (GSVI)¶
For all gamma's, both VI and GSVI yield the exact same value and policy. This is because both use the same contraction for evaluating values (but in slightly different ways which brings in other differences as explained below).
Since we do immediate value updates in GSVI (updating J(s), as and when it's new value is evaluated) instead of bulk updates like for VI (calculating all the new values H(s) and then updating J(s)=H(s)), for higher values of gamma which takes more iterations for the contraction to converge (as we have two dominant terms, whose sum we want to converge, which are 'r' (reward) and 'J' (values) in Bellman equation), we observe lesser convergence steps for GVSI as compared to VI (almost a difference of 100 iterations for gamma=1) as infered through 'extra-info' variable.
Submit to AIcrowd 🚀¶
!DATASET_PATH=$AICROWD_DATASET_PATH aicrowd notebook submit -c iit-m-rl-assignment-2-taxi -a assets
#!DATASET_PATH=$AICROWD_DATASET_PATH aicrowd notebook submit --no-verify -c iit-m-rl-assignment-2-taxi -a assets
Content
Comments
You must login before you can post a comment.