gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00
Code Issues Projects Releases Wiki Activity

small adjustment to golden tests + fixes for Doom_Health_DFP and Doom_Health_Supreme_DFP

Browse Source
This commit is contained in:
Gal Leibovich
2018-08-23 15:59:00 +03:00
parent 3fd0bf4f0f
commit 2021490caa
3 changed files with 18 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
rl_coach/presets/Doom_Health_DFP.py
View File

@@ -15,8 +15,12 @@ from rl_coach.environments.doom_environment import DoomEnvironmentParameters
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(6250000)
schedule_params.steps_between_evaluation_periods = EnvironmentSteps(62500)
schedule_params.evaluation_steps = EnvironmentSteps(6250)
# original paper evaluates according to these. But, this preset converges significantly faster - can be evaluated
# much often.
# schedule_params.steps_between_evaluation_periods = EnvironmentSteps(62500)
# schedule_params.evaluation_steps = EnvironmentSteps(6250)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(5)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
# There is no heatup for DFP. heatup length is determined according to batch size. See below.
@@ -31,6 +35,8 @@ agent_params.exploration.epsilon_schedule = LinearSchedule(0.5, 0, 10000)
agent_params.exploration.evaluation_epsilon = 0
agent_params.algorithm.goal_vector = [1] # health
# this works better than the default which is set to 8 (while running with 8 workers)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(1)
# scale observation and measurements to be -0.5 <-> 0.5
agent_params.network_wrappers['main'].input_embedders_parameters['measurements'].input_rescaling['vector'] = 100.

11
rl_coach/presets/Doom_Health_Supreme_DFP.py
View File

@@ -15,8 +15,12 @@ from rl_coach.environments.doom_environment import DoomEnvironmentParameters
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentSteps(6250000)
schedule_params.steps_between_evaluation_periods = EnvironmentSteps(62500)
schedule_params.evaluation_steps = EnvironmentSteps(6250)
# original paper evaluates according to these. But, this preset converges significantly faster - can be evaluated
# much often.
# schedule_params.steps_between_evaluation_periods = EnvironmentSteps(62500)
# schedule_params.evaluation_steps = EnvironmentSteps(6250)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(5)
schedule_params.evaluation_steps = EnvironmentEpisodes(1)
# There is no heatup for DFP. heatup length is determined according to batch size. See below.
@@ -31,6 +35,9 @@ agent_params.exploration.epsilon_schedule = LinearSchedule(0.5, 0, 10000)
agent_params.exploration.evaluation_epsilon = 0
agent_params.algorithm.goal_vector = [1] # health
# this works better than the default which is set to 8 (while running with 8 workers)
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(1)
# scale observation and measurements to be -0.5 <-> 0.5
agent_params.network_wrappers['main'].input_embedders_parameters['measurements'].input_rescaling['vector'] = 100.
agent_params.network_wrappers['main'].input_embedders_parameters['measurements'].input_offset['vector'] = 0.5

2
rl_coach/tests/golden_tests.py
View File

@@ -32,7 +32,7 @@ import time
from rl_coach.logger import screen
def read_csv_paths(test_path, filename_pattern, read_csv_tries=50):
def read_csv_paths(test_path, filename_pattern, read_csv_tries=100):
csv_paths = []
tries_counter = 0
while not csv_paths: