3
$\begingroup$

I have noticed that DDPG does rather well at solving environments with a static target.

For example, the default of Lunar Lander, the flags do not change position. So the DDPG model learns how to get to the center of the screen and land fairly quickly.

As soon as I start moving the landing position around randomly and adding the landing position as an input to the model, the model has an extremely hard time putting this connection together.

A few questions/points about adding this complexity to the environment:

  1. Would more nodes per layer or more layers help in figuring out this connection? I have tested this but seems the bigger I go, the harder it is to learn anything.
  2. Is it a common RL AI issue that it has a hard time connecting data?
  3. I realize that I could change the environment to always have a static target and instead change the position of the lunar lander ship, which in effect accomplishes the same thing, but want to know if we could solve it with moving target
  4. Is there any good documentation on Actor/Critic analyzing models? I have some results where my critic target is falling out but my critic loss is going down nicely. At the same time my actor target is going up and up and eventually plateau. It is hard to really understand what is happening and would be great to understand actor loss vs critic loss vs critic target.

Essentially, I added a random int (left side of flags), added 1 to get x position middle of landing and add 1 more to get distance of flags to be 3 out of 11 chunks.

rand_chunk = random.randint(0, CHUNKS-3)
self.x_pos_middle_landing = chunk_x[rand_chunk + 1]
self.helipad_x1 = chunk_x[rand_chunk]
self.helipad_x2 = chunk_x[rand_chunk + 2]
height[rand_chunk] = self.helipad_y
height[rand_chunk + 1] = self.helipad_y
height[rand_chunk + 2] = self.helipad_y

Old State:

state = [
        (pos.x - VIEWPORT_W/SCALE/2) / (VIEWPORT_W/SCALE/2),
        (pos.y - (self.helipad_y+LEG_DOWN/SCALE)) / (VIEWPORT_H/SCALE/2),
        vel.x*(VIEWPORT_W/SCALE/2)/FPS,
        vel.y*(VIEWPORT_H/SCALE/2)/FPS,
        self.lander.angle,
        20.0*self.lander.angularVelocity/FPS,
        1.0 if self.legs[0].ground_contact else 0.0,
        1.0 if self.legs[1].ground_contact else 0.0
        ]

Added to State, same as state[0] but using middle of 3 for landing

(self.x_pos_middle_landing - VIEWPORT_W/SCALE/2) / (VIEWPORT_W/SCALE/2)

And update the obs space from 8 spaces to 9

self.observation_space = spaces.Box(-np.inf, np.inf, shape=(9,), dtype=np.float32)

Rewards need to be updated, Old Rewards:

shaping = \
        - 100*np.sqrt(state[0]*state[0] + state[1]*state[1]) \
        - 100*np.sqrt(state[2]*state[2] + state[3]*state[3]) \
        - 100*abs(state[4]) + 10*state[6] + 10*state[7]  # And ten points for legs contact, the idea is if you

New Rewards:

shaping = \
        - 100*np.sqrt((state[0]-state[8])*(state[0]-state[8]) + state[1]*state[1]) \
        - 100*np.sqrt(state[2]*state[2] + state[3]*state[3]) \
        - 100*abs(state[4]) + 10*state[6] + 10*state[7]  # And ten points for legs contact, the idea is if you

DDPG Model

    import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
import gym
from tensorflow.keras.models import load_model
import os
import envs
import time
import scipy.stats as stats
# from stable_baselines.common.policies import MlpPolicy, MlpLstmPolicy
from stable_baselines.common.vec_env import SubprocVecEnv, DummyVecEnv, VecCheckNan
from stable_baselines.common import set_global_seeds, make_vec_env
from stable_baselines.ddpg import DDPG
from stable_baselines.ddpg.policies import MlpPolicy
# from stable_baselines.sac.policies import MlpPolicy
from stable_baselines import PPO2, SAC
from stable_baselines.common.noise import NormalActionNoise, OrnsteinUhlenbeckActionNoise, AdaptiveParamNoiseSpec

if __name__ == '__main__':
    num_cpu = 1  # Number of processes to use
    env = SubprocVecEnv([make_env(env_id, i) for i in range(num_cpu)])
    env = VecCheckNan(env, raise_exception=True, check_inf=True)
    n_actions = env.action_space.shape[-1]

    #### DDPG
    policy_kwargs = dict(act_fun=tf.nn.sigmoid, layers=[512, 512, 512], layer_norm=False)
    param_noise = None
    # action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(0.5) * np.ones(n_actions))
    action_noise = NormalActionNoise(0, 0.1)
    model = DDPG

    # # Train Model
    model.learn(total_timesteps=int(3e5))
    model.save('./models/lunar_lander')

Full Code:

    """
Rocket trajectory optimization is a classic topic in Optimal Control.

According to Pontryagin's maximum principle it's optimal to fire engine full throttle or
turn it off. That's the reason this environment is OK to have discreet actions (engine on or off).

The landing pad is always at coordinates (0,0). The coordinates are the first two numbers in the state vector.
Reward for moving from the top of the screen to the landing pad and zero speed is about 100..140 points.
If the lander moves away from the landing pad it loses reward. The episode finishes if the lander crashes or
comes to rest, receiving an additional -100 or +100 points. Each leg with ground contact is +10 points.
Firing the main engine is -0.3 points each frame. Firing the side engine is -0.03 points each frame.
Solved is 200 points.

Landing outside the landing pad is possible. Fuel is infinite, so an agent can learn to fly and then land
on its first attempt. Please see the source code for details.

To see a heuristic landing, run:

python gym/envs/box2d/lunar_lander.py

To play yourself, run:

python examples/agents/keyboard_agent.py LunarLander-v2

Created by Oleg Klimov. Licensed on the same terms as the rest of OpenAI Gym.
"""


import sys, math
import numpy as np
import random

import Box2D
from Box2D.b2 import (edgeShape, circleShape, fixtureDef, polygonShape, revoluteJointDef, contactListener)

import gym
from gym import spaces
from gym.utils import seeding, EzPickle

FPS = 50
SCALE = 30.0   # affects how fast-paced the game is, forces should be adjusted as well

MAIN_ENGINE_POWER = 13.0
SIDE_ENGINE_POWER = 0.6

INITIAL_RANDOM = 1000.0   # Set 1500 to make game harder

LANDER_POLY =[
    (-14, +17), (-17, 0), (-17 ,-10),
    (+17, -10), (+17, 0), (+14, +17)
    ]
LEG_AWAY = 20
LEG_DOWN = 18
LEG_W, LEG_H = 2, 8
LEG_SPRING_TORQUE = 40

SIDE_ENGINE_HEIGHT = 14.0
SIDE_ENGINE_AWAY = 12.0

VIEWPORT_W = 600
VIEWPORT_H = 400


class ContactDetector(contactListener):
    def __init__(self, env):
        contactListener.__init__(self)
        self.env = env

    def BeginContact(self, contact):
        if self.env.lander == contact.fixtureA.body or self.env.lander == contact.fixtureB.body:
            self.env.game_over = True
        for i in range(2):
            if self.env.legs[i] in [contact.fixtureA.body, contact.fixtureB.body]:
                self.env.legs[i].ground_contact = True

    def EndContact(self, contact):
        for i in range(2):
            if self.env.legs[i] in [contact.fixtureA.body, contact.fixtureB.body]:
                self.env.legs[i].ground_contact = False


class LunarLander(gym.Env, EzPickle):
    metadata = {
        'render.modes': ['human', 'rgb_array'],
        'video.frames_per_second' : FPS
    }

    continuous = False

    def __init__(self):
        EzPickle.__init__(self)
        self.seed()
        self.viewer = None

        self.world = Box2D.b2World()
        self.moon = None
        self.lander = None
        self.particles = []

        self.prev_reward = None

        # useful range is -1 .. +1, but spikes can be higher
        self.observation_space = spaces.Box(-np.inf, np.inf, shape=(9,), dtype=np.float32)

        if self.continuous:
            # Action is two floats [main engine, left-right engines].
            # Main engine: -1..0 off, 0..+1 throttle from 50% to 100% power. Engine can't work with less than 50% power.
            # Left-right:  -1.0..-0.5 fire left engine, +0.5..+1.0 fire right engine, -0.5..0.5 off
            self.action_space = spaces.Box(-1, +1, (2,), dtype=np.float32)
        else:
            # Nop, fire left engine, main engine, right engine
            self.action_space = spaces.Discrete(4)

        self.reset()

    def seed(self, seed=None):
        self.np_random, seed = seeding.np_random(seed)
        return [seed]

    def _destroy(self):
        if not self.moon: return
        self.world.contactListener = None
        self._clean_particles(True)
        self.world.DestroyBody(self.moon)
        self.moon = None
        self.world.DestroyBody(self.lander)
        self.lander = None
        self.world.DestroyBody(self.legs[0])
        self.world.DestroyBody(self.legs[1])

    def reset(self):
        self._destroy()
        self.world.contactListener_keepref = ContactDetector(self)
        self.world.contactListener = self.world.contactListener_keepref
        self.game_over = False
        self.prev_shaping = None

        W = VIEWPORT_W/SCALE
        H = VIEWPORT_H/SCALE

        # terrain
        CHUNKS = 11
        height = self.np_random.uniform(0, H/2, size=(CHUNKS+1,))
        chunk_x = [W/(CHUNKS-1)*i for i in range(CHUNKS)]
        rand_chunk = random.randint(0, CHUNKS-3)
        self.x_pos_middle_landing = chunk_x[rand_chunk + 1]
        self.helipad_x1 = chunk_x[rand_chunk]
        self.helipad_x2 = chunk_x[rand_chunk + 2]
        self.helipad_y = H/4
        height[rand_chunk] = self.helipad_y
        height[rand_chunk + 1] = self.helipad_y
        height[rand_chunk + 2] = self.helipad_y

        self.moon = self.world.CreateStaticBody(shapes=edgeShape(vertices=[(0, 0), (W, 0)]))
        self.sky_polys = []
        for i in range(CHUNKS-1):
            p1 = (chunk_x[i], height[i])
            p2 = (chunk_x[i+1], height[i+1])
            self.moon.CreateEdgeFixture(
                vertices=[p1,p2],
                density=0,
                friction=0.1)
            self.sky_polys.append([p1, p2, (p2[0], H), (p1[0], H)])

        self.moon.color1 = (0.0, 0.0, 0.0)
        self.moon.color2 = (0.0, 0.0, 0.0)

        initial_y = VIEWPORT_H/SCALE
        self.lander = self.world.CreateDynamicBody(
            position=(VIEWPORT_W/SCALE/2, initial_y),
            angle=0.0,
            fixtures = fixtureDef(
                shape=polygonShape(vertices=[(x/SCALE, y/SCALE) for x, y in LANDER_POLY]),
                density=5.0,
                friction=0.1,
                categoryBits=0x0010,
                maskBits=0x001,   # collide only with ground
                restitution=0.0)  # 0.99 bouncy
                )
        self.lander.color1 = (0.5, 0.4, 0.9)
        self.lander.color2 = (0.3, 0.3, 0.5)
        self.lander.ApplyForceToCenter( (
            self.np_random.uniform(-INITIAL_RANDOM, INITIAL_RANDOM),
            self.np_random.uniform(-INITIAL_RANDOM, INITIAL_RANDOM)
            ), True)

        self.legs = []
        for i in [-1, +1]:
            leg = self.world.CreateDynamicBody(
                position=(VIEWPORT_W/SCALE/2 - i*LEG_AWAY/SCALE, initial_y),
                angle=(i * 0.05),
                fixtures=fixtureDef(
                    shape=polygonShape(box=(LEG_W/SCALE, LEG_H/SCALE)),
                    density=1.0,
                    restitution=0.0,
                    categoryBits=0x0020,
                    maskBits=0x001)
                )
            leg.ground_contact = False
            leg.color1 = (0.5, 0.4, 0.9)
            leg.color2 = (0.3, 0.3, 0.5)
            rjd = revoluteJointDef(
                bodyA=self.lander,
                bodyB=leg,
                localAnchorA=(0, 0),
                localAnchorB=(i * LEG_AWAY/SCALE, LEG_DOWN/SCALE),
                enableMotor=True,
                enableLimit=True,
                maxMotorTorque=LEG_SPRING_TORQUE,
                motorSpeed=+0.3 * i  # low enough not to jump back into the sky
                )
            if i == -1:
                rjd.lowerAngle = +0.9 - 0.5  # The most esoteric numbers here, angled legs have freedom to travel within
                rjd.upperAngle = +0.9
            else:
                rjd.lowerAngle = -0.9
                rjd.upperAngle = -0.9 + 0.5
            leg.joint = self.world.CreateJoint(rjd)
            self.legs.append(leg)

        self.drawlist = [self.lander] + self.legs

        return self.step(np.array([0, 0]) if self.continuous else 0)[0]

    def _create_particle(self, mass, x, y, ttl):
        p = self.world.CreateDynamicBody(
            position = (x, y),
            angle=0.0,
            fixtures = fixtureDef(
                shape=circleShape(radius=2/SCALE, pos=(0, 0)),
                density=mass,
                friction=0.1,
                categoryBits=0x0100,
                maskBits=0x001,  # collide only with ground
                restitution=0.3)
                )
        p.ttl = ttl
        self.particles.append(p)
        self._clean_particles(False)
        return p

    def _clean_particles(self, all):
        while self.particles and (all or self.particles[0].ttl < 0):
            self.world.DestroyBody(self.particles.pop(0))

    def step(self, action):
        if self.continuous:
            action = np.clip(action, -1, +1).astype(np.float32)
        else:
            assert self.action_space.contains(action), "%r (%s) invalid " % (action, type(action))

        # Engines
        tip  = (math.sin(self.lander.angle), math.cos(self.lander.angle))
        side = (-tip[1], tip[0])
        dispersion = [self.np_random.uniform(-1.0, +1.0) / SCALE for _ in range(2)]

        m_power = 0.0
        if (self.continuous and action[0] > 0.0) or (not self.continuous and action == 2):
            # Main engine
            if self.continuous:
                m_power = (np.clip(action[0], 0.0,1.0) + 1.0)*0.5   # 0.5..1.0
                assert m_power >= 0.5 and m_power <= 1.0
            else:
                m_power = 1.0
            ox = (tip[0] * (4/SCALE + 2 * dispersion[0]) +
                side[0] * dispersion[1])  # 4 is move a bit downwards, +-2 for randomness
            oy = -tip[1] * (4/SCALE + 2 * dispersion[0]) - side[1] * dispersion[1]
            impulse_pos = (self.lander.position[0] + ox, self.lander.position[1] + oy)
            p = self._create_particle(3.5,  # 3.5 is here to make particle speed adequate
                                    impulse_pos[0],
                                    impulse_pos[1],
                                    m_power)  # particles are just a decoration
            p.ApplyLinearImpulse((ox * MAIN_ENGINE_POWER * m_power, oy * MAIN_ENGINE_POWER * m_power),
                                impulse_pos,
                                True)
            self.lander.ApplyLinearImpulse((-ox * MAIN_ENGINE_POWER * m_power, -oy * MAIN_ENGINE_POWER * m_power),
                                        impulse_pos,
                                        True)

        s_power = 0.0
        if (self.continuous and np.abs(action[1]) > 0.5) or (not self.continuous and action in [1, 3]):
            # Orientation engines
            if self.continuous:
                direction = np.sign(action[1])
                s_power = np.clip(np.abs(action[1]), 0.5, 1.0)
                assert s_power >= 0.5 and s_power <= 1.0
            else:
                direction = action-2
                s_power = 1.0
            ox = tip[0] * dispersion[0] + side[0] * (3 * dispersion[1] + direction * SIDE_ENGINE_AWAY/SCALE)
            oy = -tip[1] * dispersion[0] - side[1] * (3 * dispersion[1] + direction * SIDE_ENGINE_AWAY/SCALE)
            impulse_pos = (self.lander.position[0] + ox - tip[0] * 17/SCALE,
                        self.lander.position[1] + oy + tip[1] * SIDE_ENGINE_HEIGHT/SCALE)
            p = self._create_particle(0.7, impulse_pos[0], impulse_pos[1], s_power)
            p.ApplyLinearImpulse((ox * SIDE_ENGINE_POWER * s_power, oy * SIDE_ENGINE_POWER * s_power),
                                impulse_pos
                                , True)
            self.lander.ApplyLinearImpulse((-ox * SIDE_ENGINE_POWER * s_power, -oy * SIDE_ENGINE_POWER * s_power),
                                        impulse_pos,
                                        True)

        self.world.Step(1.0/FPS, 6*30, 2*30)

        pos = self.lander.position
        vel = self.lander.linearVelocity
        state = [
            (pos.x - VIEWPORT_W/SCALE/2) / (VIEWPORT_W/SCALE/2),
            (pos.y - (self.helipad_y+LEG_DOWN/SCALE)) / (VIEWPORT_H/SCALE/2),
            vel.x*(VIEWPORT_W/SCALE/2)/FPS,
            vel.y*(VIEWPORT_H/SCALE/2)/FPS,
            self.lander.angle,
            20.0*self.lander.angularVelocity/FPS,
            1.0 if self.legs[0].ground_contact else 0.0,
            1.0 if self.legs[1].ground_contact else 0.0,
            (self.x_pos_middle_landing - VIEWPORT_W/SCALE/2) / (VIEWPORT_W/SCALE/2)
            ]
        assert len(state) == 9

        reward = 0
        shaping = \
            - 100*np.sqrt((state[0]-state[8])*(state[0]-state[8]) + state[1]*state[1]) \
            - 100*np.sqrt(state[2]*state[2] + state[3]*state[3]) \
            - 100*abs(state[4]) + 10*state[6] + 10*state[7]  # And ten points for legs contact, the idea is if you
                                                            # lose contact again after landing, you get negative reward
        if self.prev_shaping is not None:
            reward = shaping - self.prev_shaping
        self.prev_shaping = shaping

        reward -= m_power*0.30  # less fuel spent is better, about -30 for heuristic landing
        reward -= s_power*0.03

        done = False
        if self.game_over or abs(state[0]) >= 1.0:
            done = True
            reward = -100
        if not self.lander.awake:
            done = True
            reward = +100
        return np.array(state, dtype=np.float32), reward, done, {}

    def render(self, mode='human'):
        from gym.envs.classic_control import rendering
        if self.viewer is None:
            self.viewer = rendering.Viewer(VIEWPORT_W, VIEWPORT_H)
            self.viewer.set_bounds(0, VIEWPORT_W/SCALE, 0, VIEWPORT_H/SCALE)

        for obj in self.particles:
            obj.ttl -= 0.15
            obj.color1 = (max(0.2, 0.2+obj.ttl), max(0.2, 0.5*obj.ttl), max(0.2, 0.5*obj.ttl))
            obj.color2 = (max(0.2, 0.2+obj.ttl), max(0.2, 0.5*obj.ttl), max(0.2, 0.5*obj.ttl))

        self._clean_particles(False)

        for p in self.sky_polys:
            self.viewer.draw_polygon(p, color=(0, 0, 0))

        for obj in self.particles + self.drawlist:
            for f in obj.fixtures:
                trans = f.body.transform
                if type(f.shape) is circleShape:
                    t = rendering.Transform(translation=trans*f.shape.pos)
                    self.viewer.draw_circle(f.shape.radius, 20, color=obj.color1).add_attr(t)
                    self.viewer.draw_circle(f.shape.radius, 20, color=obj.color2, filled=False, linewidth=2).add_attr(t)
                else:
                    path = [trans*v for v in f.shape.vertices]
                    self.viewer.draw_polygon(path, color=obj.color1)
                    path.append(path[0])
                    self.viewer.draw_polyline(path, color=obj.color2, linewidth=2)

        for x in [self.helipad_x1, self.helipad_x2]:
            flagy1 = self.helipad_y
            flagy2 = flagy1 + 50/SCALE
            self.viewer.draw_polyline([(x, flagy1), (x, flagy2)], color=(1, 1, 1))
            self.viewer.draw_polygon([(x, flagy2), (x, flagy2-10/SCALE), (x + 25/SCALE, flagy2 - 5/SCALE)],
                                    color=(0.8, 0.8, 0))

        return self.viewer.render(return_rgb_array=mode == 'rgb_array')

    def close(self):
        if self.viewer is not None:
            self.viewer.close()
            self.viewer = None


class RandomTargetLunarLander(LunarLander):
    continuous = True

def heuristic(env, s):
    """
    The heuristic for
    1. Testing
    2. Demonstration rollout.

    Args:
        env: The environment
        s (list): The state. Attributes:
                s[0] is the horizontal coordinate
                s[1] is the vertical coordinate
                s[2] is the horizontal speed
                s[3] is the vertical speed
                s[4] is the angle
                s[5] is the angular speed
                s[6] 1 if first leg has contact, else 0
                s[7] 1 if second leg has contact, else 0
                s[8] is the target coordinate
    returns:
        a: The heuristic to be fed into the step function defined above to determine the next step and reward.
    """

    angle_targ = s[0]*0.5 + s[2]*1.0         # angle should point towards center
    if angle_targ > 0.4: angle_targ = 0.4    # more than 0.4 radians (22 degrees) is bad
    if angle_targ < -0.4: angle_targ = -0.4
    hover_targ = 0.55*np.abs(s[0])           # target y should be proportional to horizontal offset

    angle_todo = (angle_targ - s[4]) * 0.5 - (s[5])*1.0
    hover_todo = (hover_targ - s[1])*0.5 - (s[3])*0.5

    if s[6] or s[7]:  # legs have contact
        angle_todo = 0
        hover_todo = -(s[3])*0.5  # override to reduce fall speed, that's all we need after contact

    if env.continuous:
        a = np.array([hover_todo*20 - 1, -angle_todo*20])
        a = np.clip(a, -1, +1)
    else:
        a = 0
        if hover_todo > np.abs(angle_todo) and hover_todo > 0.05: a = 2
        elif angle_todo < -0.05: a = 3
        elif angle_todo > +0.05: a = 1
    return a

def demo_heuristic_lander(env, seed=None, render=False):
    env.seed(seed)
    total_reward = 0
    steps = 0
    s = env.reset()
    while True:
        a = heuristic(env, s)
        s, r, done, info = env.step(a)
        total_reward += r

        if render:
            still_open = env.render()
            if still_open == False: break

        if steps % 20 == 0 or done:
            print("observations:", " ".join(["{:+0.2f}".format(x) for x in s]))
            print("step {} total_reward {:+0.2f}".format(steps, total_reward))
        steps += 1
        if done: break
    return total_reward


if __name__ == '__main__':
    demo_heuristic_lander(LunarLander(), render=True)

Here are the results: Green is Normal Lunar Lander Continuous Pink is the Random Target Lunar Lander Continuous

Results_1 Results_2 Results_3

$\endgroup$
3
  • $\begingroup$ Quick question before going any deeper - how have you scaled the new flag position values for input to the NN - i.e. what range can they take "naturally" in the environment, and what values as input to the NN do you use to represent those natural values? $\endgroup$ – Neil Slater May 3 at 8:59
  • $\begingroup$ I have changed the title of this post to put what I think is your main question. Feel free to edit your post again if that's not the case. $\endgroup$ – nbro May 3 at 12:30
  • $\begingroup$ Will post code shortly $\endgroup$ – user1779362 May 3 at 15:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.