IIT-M RL-ASSIGNMENT-2-GRIDWORLD
Solution for submission 132309
A detailed solution for submission 132309 submitted for challenge IIT-M RL-ASSIGNMENT-2-GRIDWORLD
What is the notebook about?¶
Problem - Gridworld Environment Algorithms¶
This problem deals with a grid world and stochastic actions. The tasks you have to do are:
- Implement Policy Iteration
- Implement Value Iteration
- Implement TD lamdda
- 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 π 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
AICROWD_DATASET_PATH = os.getenv("DATASET_PATH", os.getcwd()+"/a5562c7d-55f0-4d06-841c-110655bb04ec_a2_gridworld_inputs.zip")
AICROWD_RESULTS_DIR = os.getenv("OUTPUTS_DIR", "results")
!unzip -q $AICROWD_DATASET_PATH
DATASET_DIR = 'inputs/'
GridWorld Environment¶
Read the code for the environment thoroughly
Do not edit the code for the environment
import numpy as np
class GridEnv_HW2:
def __init__(self,
goal_location,
action_stochasticity,
non_terminal_reward,
terminal_reward,
grey_in,
brown_in,
grey_out,
brown_out
):
# Do not edit this section
self.action_stochasticity = action_stochasticity
self.non_terminal_reward = non_terminal_reward
self.terminal_reward = terminal_reward
self.grid_size = [10, 10]
# Index of the actions
self.actions = {'N': (1, 0),
'E': (0,1),
'S': (-1,0),
'W': (0,-1)}
self.perpendicular_order = ['N', 'E', 'S', 'W']
l = ['normal' for _ in range(self.grid_size[0]) ]
self.grid = np.array([l for _ in range(self.grid_size[1]) ], dtype=object)
self.grid[goal_location[0], goal_location[1]] = 'goal'
self.goal_location = goal_location
for gi in grey_in:
self.grid[gi[0],gi[1]] = 'grey_in'
for bi in brown_in:
self.grid[bi[0], bi[1]] = 'brown_in'
for go in grey_out:
self.grid[go[0], go[1]] = 'grey_out'
for bo in brown_out:
self.grid[bo[0], bo[1]] = 'brown_out'
self.grey_outs = grey_out
self.brown_outs = brown_out
def _out_of_grid(self, state):
if state[0] < 0 or state[1] < 0:
return True
elif state[0] > self.grid_size[0] - 1:
return True
elif state[1] > self.grid_size[1] - 1:
return True
else:
return False
def _grid_state(self, state):
return self.grid[state[0], state[1]]
def get_transition_probabilites_and_reward(self, state, action):
"""
Returns the probabiltity of all possible transitions for the given action in the form:
A list of tuples of (next_state, probability, reward)
Note that based on number of state and action there can be many different next states
Unless the state is All the probabilities of next states should add up to 1
"""
grid_state = self._grid_state(state)
if grid_state == 'goal':
return [(self.goal_location, 1.0, 0.0)]
elif grid_state == 'grey_in':
npr = []
for go in self.grey_outs:
npr.append((go, 1/len(self.grey_outs),
self.non_terminal_reward))
return npr
elif grid_state == 'brown_in':
npr = []
for bo in self.brown_outs:
npr.append((bo, 1/len(self.brown_outs),
self.non_terminal_reward))
return npr
direction = self.actions.get(action, None)
if direction is None:
raise ValueError("Invalid action %s , please select among" % action, list(self.actions.keys()))
dir_index = self.perpendicular_order.index(action)
wrap_acts = self.perpendicular_order[dir_index:] + self.perpendicular_order[:dir_index]
next_state_probs = {}
for prob, a in zip(self.action_stochasticity, wrap_acts):
d = self.actions[a]
next_state = (state[0] + d[0]), (state[1] + d[1])
if self._out_of_grid(next_state):
next_state = state
next_state_probs.setdefault(next_state, 0.0)
next_state_probs[next_state] += prob
npr = []
for ns, prob in next_state_probs.items():
next_grid_state = self._grid_state(ns)
reward = self.terminal_reward if next_grid_state == 'goal' else self.non_terminal_reward
npr.append((ns, prob, reward))
return npr
def step(self, state, action):
npr = self.get_transition_probabilites_and_reward(state, action)
probs = [t[1] for t in npr]
sampled_idx = np.random.choice(range(len(npr)), p=probs)
sampled_npr = npr[sampled_idx]
next_state = sampled_npr[0]
reward = sampled_npr[2]
is_terminal = next_state == tuple(self.goal_location)
return next_state, reward, is_terminal
Example environment¶
This has the same setup as the pdf, do not edit the settings
def get_base_kwargs():
goal_location = (9,9)
action_stochasticity = [0.8, 0.2/3, 0.2/3, 0.2/3]
grey_out = [(3,2), (4,2), (5,2), (6,2)]
brown_in = [(9,7)]
grey_in = [(0,0)]
brown_out = [(1,7)]
non_terminal_reward = 0
terminal_reward = 10
base_kwargs = {"goal_location": goal_location,
"action_stochasticity": action_stochasticity,
"brown_in": brown_in,
"grey_in": grey_in,
"brown_out": brown_out,
"non_terminal_reward": non_terminal_reward,
"terminal_reward": terminal_reward,
"grey_out": grey_out,}
return base_kwargs
base_kwargs = get_base_kwargs()
Task 2.1 - Value Iteration¶
Run value iteration on the environment and generate the policy and expected reward
def value_iteration(env, gamma):
# Initial Values
values = np.zeros((10, 10))
# Initial policy
policy = np.empty((10, 10), object)
policy[:] = 'N' # Make all the policy values as 'N'
# Begin code here
iteration_states={}
visited=[]
H=np.zeros((10,10))
iteration=0
while(True):
iteration=iteration+1
delta = 0
prev_grid = values.copy()
for i in range(10):
for j in range(10):
#value = prev_grid[i][j]
for action_no,action in enumerate(env.actions):
temp_value = 0
possible_transitions = env.get_transition_probabilites_and_reward((i,j),action)
for transitions in possible_transitions:
temp_value = temp_value + (transitions[2] + gamma*prev_grid[transitions[0][0]][transitions[0][1]])*transitions[1]
if(action_no==0):
H[i][j] = temp_value
policy[i][j]=action
else:
if(H[i][j]<temp_value):
policy[i][j] = action
H[i][j] = temp_value
delta = max(delta,abs(H[i][j]-values[i][j]))
if(abs(H[i][j]-values[i][j])<1e-8):
if(env._grid_state((i,j))=='grey_in' and ((i,j) not in visited)):
iteration_states["grey_in"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='brown_in' and ((i,j) not in visited)):
iteration_states["brown_in"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='brown_out' and ((i,j) not in visited)):
iteration_states["brown_out"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='grey_out' and ((i,j) not in visited)):
iteration_states["grey_out"]=iteration
visited.append((i,j))
for i in range(10):
for j in range(10):
values[i][j]=H[i][j]
if(delta<1e-8):
break
# Put your extra information needed for plots etc in this dictionary
extra_info = {}
extra_info=iteration_states
# End code
# Do not change the number of output values
return {"Values": values, "Policy": policy}, extra_info
env = GridEnv_HW2(**base_kwargs)
res, extra_info = value_iteration(env, 0.7)
# The rounding off is just for making print statement cleaner
print(np.flipud(np.round(res['Values'], decimals=2)))
print(np.flipud(res['Policy']))
Task 2.2 - Policy Iteration¶
Run policy iteration on the environment and generate the policy and expected reward
def policy_iteration(env, gamma):
# Initial Values
values = np.zeros((10, 10))
# Initial policy
policy = np.empty((10, 10), object)
policy[:] = 'N' # Make all the policy values as 'N'
# Begin code here
iteration_states={}
visited=[]
iteration=0
values_iteration=[]
while(True):
iteration=iteration+1
while(True):
prev_grid=values.copy()
delta=0
for i in range(10):
for j in range(10):
value_temp = values[i][j]
action=policy[i][j]
possible_transitions = env.get_transition_probabilites_and_reward((i,j),action)
temp_value=0
for transitions in possible_transitions:
temp_value = temp_value + (transitions[2] + gamma*prev_grid[transitions[0][0]][transitions[0][1]])*transitions[1]
values[i][j]=temp_value
delta = max(delta,abs(temp_value-value_temp))
if(delta<1e-8):
break
values_iteration.append(values.copy())
done=1
for i in range(10):
for j in range(10):
b=policy[i][j]
best=0
for action_no,action in enumerate(env.actions):
temp_value = 0
possible_transitions = env.get_transition_probabilites_and_reward((i,j),action)
for transitions in possible_transitions:
temp_value = temp_value + (transitions[2] + gamma*prev_grid[transitions[0][0]][transitions[0][1]])*transitions[1]
if(action_no==0):
best=temp_value
policy[i][j]=action
else:
if(best<temp_value):
policy[i][j] = action
best=temp_value
if(b!=policy[i][j]):
done=0
if(b==policy[i][j]):
if(env._grid_state((i,j))=='grey_in' and ((i,j) not in visited)):
iteration_states["grey_in"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='brown_in' and ((i,j) not in visited)):
iteration_states["brown_in"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='brown_out' and ((i,j) not in visited)):
iteration_states["brown_out"]=iteration
visited.append((i,j))
elif(env._grid_state((i,j))=='grey_out' and ((i,j) not in visited)):
iteration_states["grey_out"]=iteration
visited.append((i,j))
if(done==1):
break
error_plot=[]
for value_iter in values_iteration:
error=0
for i in range(10):
for j in range(10):
error = error + (value_iter[i][j]-values[i][j])*(value_iter[i][j]-values[i][j])
error=error/100
error_plot.append(error)
# Put your extra information needed for plots etc in this dictionary
extra_info = {}
extra_info["iterations"]=iteration_states
extra_info["errors"]=error_plot
# End code
# Do not change the number of output values
return {"Values": values, "Policy": policy}, extra_info
env = GridEnv_HW2(**base_kwargs)
res, extra_info = policy_iteration(env, 0.7)
# The rounding off is just for making print statement cleaner
print(np.flipud(np.round(res['Values'], decimals=2)))
print(np.flipud(res['Policy']))
Task 2.3 - TD Lambda¶
Use the heuristic policy and implement TD lambda to find values on the gridworld
# The policy mentioned in the pdf to be used for TD lambda, do not modify this
def heuristic_policy(env, state):
goal = env.goal_location
dx = goal[0] - state[0]
dy = goal[1] - state[1]
if abs(dx) >= abs(dy):
direction = (np.sign(dx), 0)
else:
direction = (0, np.sign(dy))
for action, dir_val in env.actions.items():
if dir_val == direction:
target_action = action
break
return target_action
def td_lambda(env, lamda, seeds):
alpha = 0.5
gamma = 0.7
N = len(seeds)
# Usage of input_policy
# heuristic_policy(env, state) -> action
example_action = heuristic_policy(env, (1,2)) # Returns 'N' if goal is (9,9)
# Example of env.step
# env.step(state, action) -> Returns next_state, reward, is_terminal
# Initial values
values = np.zeros((10, 10))
es = np.zeros((10,10))
iterations=[]
values_iteration=[]
for episode_idx in range(N):
# Do not change this else the results will not match due to environment stochas
np.random.seed(seeds[episode_idx])
grey_in_loc = np.where(env.grid == 'grey_in')
state = grey_in_loc[0][0], grey_in_loc[1][0]
done = False
iteration=0
while not done:
action = heuristic_policy(env, state)
ns, rew, is_terminal = env.step(state, action)
# env.step is already taken inside the loop for you,
# Don't use env.step anywhere else in your code
# Begin code here
delta=rew+(gamma*values[ns[0]][ns[1]])-values[state[0]][state[1]]
es[state[0]][state[1]]=es[state[0]][state[1]]+1
for i in range(10):
for j in range(10):
values[state[0]][state[1]]=values[state[0]][state[1]]+(alpha*delta*es[state[0]][state[1]])
es[state[0]][state[1]] = gamma*lamda*es[state[0]][state[1]]
state=ns
iteration=iteration+1
if(env._grid_state(state)=='goal'):
break
values_iteration.append(values.copy())
iterations.append(iteration)
error_plot=[]
for value_iter in values_iteration:
error=0
for i in range(10):
for j in range(10):
error = error + (value_iter[i][j]-values[i][j])*(value_iter[i][j]-values[i][j])
error=error/100
error_plot.append(error)
# Put your extra information needed for plots etc in this dictionary
extra_info = {}
extra_info["iterations"]=iterations
extra_info["error"]=error_plot
# End code
# Do not change the number of output values
return {"Values": values}, extra_info
env = GridEnv_HW2(**base_kwargs)
res, extra_info = td_lambda(env, lamda=0.5, seeds=np.arange(1000))
# The rounding off is just for making print statement cleaner
print(np.flipud(np.round(res['Values'], decimals=2)))
Task 2.4 - TD Lamdba for multiple values of $\lambda$¶
Ideally this code should run as is
# This cell is only for your subjective evaluation results, display the results as asked in the pdf
# You can change it as you require, this code should run TD lamdba by default for different values of lambda
lamda_values = np.arange(0, 100+5, 5)/100
td_lamda_results = {}
extra_info = {}
for lamda in lamda_values:
env = GridEnv_HW2(**base_kwargs)
td_lamda_results[lamda], extra_info[lamda] = td_lambda(env, lamda,
seeds=np.arange(1000))
import matplotlib.pyplot as plt
lamdas = [0,0.25,0.5,0.75,1]
for lamda in lamdas:
plt.plot(extra_info[lamda]["error"])
plt.xlabel("Iterations")
plt.ylabel("Error Values")
plt.title(lamda)
plt.show()
Generate Results ✅¶
def get_results(kwargs):
gridenv = GridEnv_HW2(**kwargs)
policy_iteration_results = policy_iteration(gridenv, 0.7)[0]
value_iteration_results = value_iteration(gridenv, 0.7)[0]
td_lambda_results = td_lambda(env, 0.5, np.arange(1000))[0]
final_results = {}
final_results["policy_iteration"] = policy_iteration_results
final_results["value_iteration"] = value_iteration_results
final_results["td_lambda"] = td_lambda_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 score on the public data¶
This scores 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):
if 'Policy' in results:
policy_match = results['Policy'] == targets['Policy']
else:
policy_match = True
# Reference https://numpy.org/doc/stable/reference/generated/numpy.allclose.html
rewards_match = np.allclose(results['Values'], targets['Values'], rtol=3)
equal = rewards_match and policy_match
return equal
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))
Display Results of TD lambda¶
Display Results of TD lambda with lambda values from 0 to 1 with steps of 0.05
Add code/text as required
for lamda in td_lamda_results:
numbers = ['1','2','3','4','5','6','7','8','9','10']
print('Values Grid for lamda :',lamda)
plt.rcParams['figure.figsize'] = [8,8]
fig, ax = plt.subplots()
caxes = ax.matshow(td_lamda_results[lamda]['Values'])
fig.colorbar(caxes)
xaxis = np.arange(len(numbers))
ax.set_xticks(xaxis)
ax.set_yticks(xaxis)
ax.set_xticklabels(numbers)
ax.set_yticklabels(numbers)
if lamda>=0.95:
for i in range(10):
for j in range(10):
c = "{:.0e}".format(td_lamda_results[lamda]['Values'][j,i])
ax.text(i, j, str(c), va='center', ha='center')
else:
for i in range(10):
for j in range(10):
c = int(td_lamda_results[lamda]['Values'][j,i])
ax.text(i, j, str(c), va='center', ha='center')
plt.show()
Subjective questions¶
2.a Value Iteration vs Policy Iteration¶
- Compare value iteration and policy iteration for states Brown in, Brown Out, Grey out and Grey In
- Which one converges faster and why
Both value and policy iteration give the same results of values and policies after convergence.
The states grey_in and brown_in converges much faster than the other states. This is because the next state is defined very clearly in the grey_in and brown_in state whereas that is not the case for grey_out and brown_out state, so the convergence is much faster in the former case than the latter case.
2.b How changing $\lambda$ affecting TD Lambda¶
As lamda increases the values of states increases drastically as can be seen from lamda=0.95,1. This is because the higher the values of lamda the more information is collected using more steps rather than going fewer steps(like in lesser values of lamda). Lamda=1 corresponds to the monte carlo policy evaluation
2.c Policy iteration error curve¶
Plot error curve of $J_i$ vs iteration $i$ for policy iteration
res, extra_info = policy_iteration(env, 0.7)
plt.plot(extra_info["errors"])
plt.xlabel("Iterations")
plt.ylabel("Error Values")
plt.title("Policy Iteration Error Curve")
plt.show()
2.d TD Lamdba error curve¶
Plot error curve of $J_i$ vs iteration $i$ for TD Lambda for $\lambda = [0, 0.25, 0.5, 0.75, 1]$
Plotted in Task 2.4
Submit to AIcrowd 🚀¶
!DATASET_PATH=$AICROWD_DATASET_PATH aicrowd notebook submit --no-verify -c iit-m-rl-assignment-2-gridworld -a assets
Content
Comments
You must login before you can post a comment.