Rainbow Algorithm on Slime Volleyball Environment

Using Rainbow implementation in Chainer RL for Slime Volleyball Environment (Pixel Observation)

Train model from scratch:

python train_rainbow.py --outdir slime --seed 612 --gpu 0 --env SlimeVolleySurvivalNoFrameskip-v0

Run pre-trained for 1000 episodes:

python train_rainbow.py --env SlimeVolleyNoFrameskip-v0 --load zoo/best/ --demo --gpu -1 --eval-n-steps 300001

Final Result:

n_episodes: 1000 mean: 0.037 median: 0.0 stdev 0.9942935278978837

Run in render mode on desktop to visualize the agent:

python train_rainbow.py --env SlimeVolleyNoFrameskip-v0 --load zoo/best/ --demo --gpu -1 --render 

 

This is a fork of the example in Chainer RL. Below is from the original README:

Rainbow

This example trains a Rainbow agent, from the following paper: Rainbow: Combining Improvements in Deep Reinforcement Learning.

Requirements

• atari_py>=0.1.1
• opencv-python

Running the Example

To run the training example:

python train_rainbow.py [options]

We have already pretrained models from this script for all the domains list in the results section. To load a pretrained model:

python train_rainbow.py --demo --load-pretrained --env BreakoutNoFrameskip-v4 --pretrained-type best --gpu -1

Useful Options

• --gpu. Specifies the GPU. If you do not have a GPU on your machine, run the example with the option --gpu -1. E.g. python train_rainbow.py --gpu -1.
• --env. Specifies the environment.
• --render. Add this option to render the states in a GUI window.
• --seed. This option specifies the random seed used.
• --outdir This option specifies the output directory to which the results are written.
• --demo. Runs an evaluation, instead of training the agent.
• --load-pretrained Loads the pretrained model. Both --load and --load-pretrained cannot be used together.
• --pretrained-type. Either best (the best intermediate network during training) or final (the final network after training).

To view the full list of options, either view the code or run the example with the --help option.

Results

These results reflect ChainerRL v0.7.0.

Results Summary
Reporting Protocol	A re-evaluation of the best intermediate agent
Number of seeds	1
Number of common domains	52
Number of domains where paper scores higher	20
Number of domains where ChainerRL scores higher	30
Number of ties between paper and ChainerRL	2

Game	ChainerRL Score	Original Reported Scores
AirRaid	6500.9	N/A
Alien	9409.1	9491.7
Amidar	3252.7	5131.2
Assault	15245.5	14198.5
Asterix	353258.5	428200.3
Asteroids	2792.3	2712.8
Atlantis	894708.5	826659.5
BankHeist	1734.8	1358.0
BattleZone	90625.0	62010.0
BeamRider	27959.5	16850.2
Berzerk	26704.2	2545.6
Bowling	67.1	30.0
Boxing	99.8	99.6
Breakout	340.8	417.5
Carnival	5530.3	N/A
Centipede	7718.1	8167.3
ChopperCommand	303480.5	16654.0
CrazyClimber	165370.0	168788.5
Defender	N/A	55105.0
DemonAttack	110028.0	111185.2
DoubleDunk	-0.1	-0.3
Enduro	2273.8	2125.9
FishingDerby	45.3	31.3
Freeway	33.7	34.0
Frostbite	10432.3	9590.5
Gopher	76662.9	70354.6
Gravitar	1819.5	1419.3
Hero	12590.5	55887.4
IceHockey	5.1	1.1
Jamesbond	31392.0	N/A
JourneyEscape	0.0	N/A
Kangaroo	14462.5	14637.5
Krull	7989.0	8741.5
KungFuMaster	22820.5	52181.0
MontezumaRevenge	4.0	384.0
MsPacman	6153.4	5380.4
NameThisGame	14035.1	13136.0
Phoenix	5169.6	108528.6
Pitfall	0.0	0.0
Pong	20.9	20.9
Pooyan	7793.1	N/A
PrivateEye	100.0	4234.0
Qbert	42481.1	33817.5
Riverraid	26114.0	N/A
RoadRunner	64306.0	62041.0
Robotank	74.4	61.4
Seaquest	4286.8	15898.9
Skiing	-9441.0	-12957.8
Solaris	7902.2	3560.3
SpaceInvaders	2838.0	18789.0
StarGunner	181192.5	127029.0
Surround	N/A	9.7
Tennis	-0.1	0.0
TimePilot	25582.0	12926.0
Tutankham	251.9	241.0
UpNDown	284465.6	N/A
Venture	1499.0	5.5
VideoPinball	492071.8	533936.5
WizardOfWor	19796.5	17862.5
YarsRevenge	80817.2	102557.0
Zaxxon	26827.5	22209.5

Evaluation Protocol

Our evaluation protocol is designed to mirror the evaluation protocol of the original paper as closely as possible, in order to offer a fair comparison of the quality of our example implementation. Specifically, the details of our evaluation (also can be found in the code) are the following:

• Evaluation Frequency: The agent is evaluated after every 1 million frames (250K timesteps). This results in a total of 200 "intermediate" evaluations.
• Evaluation Phase: The agent is evaluated for 500K frames (125K timesteps) in each intermediate evaluation.
  • Output: The output of an intermediate evaluation phase is a score representing the mean score of all completed evaluation episodes within the 125K timesteps. If there is any unfinished episode by the time the 125K timestep evaluation phase is finished, that episode is discarded.
• Intermediate Evaluation Episode:
  • Capped at 30 mins of play, or 108K frames/ 27K timesteps.
  • Each evaluation episode begins with a random number of no-ops (up to 30), where this number is chosen uniformly at random.
  • During evaluation episodes the agent uses an epsilon-greedy policy, with epsilon=0.001 (original paper does greedy evaluation because noisy networks are used)
• Reporting: For each run of our Rainbow example, we take the network weights of the best intermediate agent (i.e. the network weights that achieved the highest intermediate evaluation), and re-evaluate that agent for 200 episodes. In each of these 200 "final evaluation" episodes, the episode is terminated after 30 minutes of play (30 minutes = 1800 seconds * 60 frames-per-second / 4 frames per action = 27000 timesteps). We then output the average of these 200 episodes as the achieved score for the Rainbow agent. The reported value in the table consists of the average of 1 "final evaluation", or 1 run of this Rainbow example.

Training times

Time statistics...

Training time (in days) across all domains
Mean	12.929
Fastest Domain	11.931 (Frostbite)
Slowest Domain	13.974 (UpNDown)