A continuous action space version of A3C LSTM in pytorch plus A3G design
Training with A3G benefits training speed most when using larger models i.e using raw pixels for observations such as training in atari environments that have raw pixels for state representation
This repository includes my implementation with reinforcement learning using Asynchronous Advantage Actor-Critic (A3C) in Pytorch an algorithm from Google Deep Mind's paper "Asynchronous Methods for Deep Reinforcement Learning."
New implementation of A3C that utilizes GPU for speed increase in training. Which we can call A3G. A3G as opposed to other versions that try to utilize GPU with A3C algorithm, with A3G each agent has its own network maintained on GPU but shared model is on CPU and agent models are quickly converted to CPU to update shared model which allows updates to be frequent and fast by utilizing Hogwild Training and make updates to shared model asynchronously and without locks. This new method greatly increase training speed and models and can be see in my rla3cpytorch repo that training that use to take days to train can be trained in as fast as 10minutes for some Atari games!
This is continuous domain version of my other a3c repo. Here I show A3C can solve BipedalWalker-v2 but also the much harder BipedalWalkerHardcore-v2 version as well. "Solved" meaning to train a model capable of averaging reward over 300 for 100 consecutive episodes
Added trained model for BipedWalkerHardcore-v2
When training model it is important to limit number of worker processes to number of cpu cores available as too many processes (e.g. more than one process per cpu core available) will actually be detrimental in training speed and effectiveness
To train agent in BipedalWalker-v2 environment with 6 different worker processes: On a MacPro 2014 laptop traing typically takes 15-20mins to get to a winning solution
python main.py --workers 6 --env BipedalWalker-v2 --save-max True --model MLP --stack-frames 1
To train agent in BipedalWalkerHardcore-v2 environment with 64 different worker processes: BipedalWalkerHardcore-v2 is much harder environment compared to normal BipedalWalker On a 72 cpu AWS EC2 c5.18xlarge instance training with 64 worker processes takes up to 24hrs to get to model that could solve the environment. Using enhanced A3G design, training model takes only 4-6hrs
python main.py --workers 64 --env BipedalWalkerHardcore-v2 --save-max True --model CONV --stack-frames 4
To train agent in BipedalWalkerHardcore-v2 environment with 32 different worker processes with new A3C-GPU:
python main.py --env BipedalWalkerHardcore-v2 --workers 32 --gpu-ids 0 1 2 3 --amsgrad True --model CONV --stack-frames 4
Hit Ctrl C to end training session properly
To run a 100 episode gym evaluation with trained model
python gym_eval.py --env BipedalWalkerHardcore-v2 --num-episodes 100 --stack-frames 4 --model CONV --new-gym-eval True