from gw_collect import Gridworld
import pygame as pg
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torch.nn.functional as F
from copy import deepcopy
from vicero.algorithms.qlearning import Qlearning
#env = Gridworld(width=6, height=6, cell_size=32, agent_pos=(0, 3), food_pos=[(0, 0), (3, 3), (4, 5), (2, 0)])
env_a = Gridworld(width=4,
height=4,
cell_size=32,
agent_pos=(0, 0),
food_pos=[(0, 3), (3, 3)])
env_b = Gridworld(width=4,
height=4,
cell_size=32,
agent_pos=(0, 0),
food_pos=[(0, 3), (3, 3)])
pg.init()
screen = pg.display.set_mode(
(env_a.cell_size * env_a.width, env_a.cell_size * env_a.height))
env_a.screen = screen
clock = pg.time.Clock()
if len(durations_t) >= 100:
means = durations_t.unfold(0, 100, 1).mean(1).view(-1)
means = torch.cat((torch.zeros(99), means))
plt.plot(means.numpy(), c='green')
plt.pause(0.001)
print('Set A')
histories_a = []
env_list = []
for i in range(10):
env_list.append(Gridworld(width=4, height=4, cell_size=32, seed=(20 + i)))
#for i in [8, 0, 2, 4, 7]:
# env_list.append(Gridworld(width=4, height=4, cell_size=32, seed=i))
env = MultitaskEnvironment(env_list)
for _ in range(repetitions):
print('x')
dqn = DQN(env,
qnet=NeuralNet(64, 4).double(),
plotter=plot,
render=False,
memory_length=2000,
gamma=.99,
alpha=.001,
epsilon_start=0.1,
state = torch.from_numpy(state)
env_diffs.append(
torch.sum((rf_a.policy_net(state) -
rf_b.policy_net(state))**2).item())
print('{}/{} mean difference: {:.4f}'.format(ne + 1, iterations,
np.mean(env_diffs)))
all_mean_diffs.append(np.mean(env_diffs))
absolutely_all_diffs.append(all_mean_diffs)
return all_mean_diffs[-1]
envs = [(Gridworld(width=4,
height=4,
cell_size=32,
agent_pos=(2, 0),
food_pos=[(1, 3), (3, 3)]),
Gridworld(width=4,
height=4,
cell_size=32,
agent_pos=(0, 0),
food_pos=[(1, 3), (3, 3)]))]
for env_pair in envs:
print(env_diff(env_pair[0], env_pair[1], 50, 100))
for diff in absolutely_all_diffs:
plt.plot(diff)
#plt.savefig('test.png')
if len(durations_t) >= 100:
means = durations_t.unfold(0, 100, 1).mean(1).view(-1)
means = torch.cat((torch.zeros(99), means))
plt.plot(means.numpy(), c='green')
plt.pause(0.001)
histories_all = []
env_list = []
for i in [8, 0, 2, 4, 7, 22, 23, 24, 25, 51]:
print(i)
env = Gridworld(width=4, height=4, cell_size=32, seed=i)
history = []
for _ in range(repetitions):
dqn = DQN(env,
qnet=LinRegNet(64, 4).double(),
plotter=None,
render=False,
memory_length=2000,
gamma=.99,
alpha=.001,
epsilon_start=0.1,
plot_durations=True)
dqn.train(training_iterations, 4, plot=False)
history.append(dqn.history)
def average_environment_duration_dqn(env_set, iterations_per_env):
all_durations = []
for env in env_set:
all_durations.append('env={},mean_duration={}'.format(
env, np.mean(dqn_benchmark(env, iterations_per_env))))
print(all_durations[-1])
return all_durations
def average_environment_duration_rein(env_set, iterations_per_env):
all_durations = []
for env in env_set:
all_durations.append('env={},mean_duration={}'.format(
env, np.mean(rein_benchmark(env, iterations_per_env))))
print(all_durations[-1])
return all_durations
#env = Gridworld(width=4, height=4, cell_size=32, seed=33)
env_set = []
for i in range(10):
env_set.append(Gridworld(width=4, height=4, cell_size=32, seed=i))
exp.run(average_environment_duration_rein,
params={
'env_set': env_set,
'iterations_per_env': 10
},
k=1)