stevenson0421/pome-implementation

self implementation of RL method Policy Optimization with Model-based Explorations

POME-implementation

self implementation of RL method Policy Optimization with Model-based Explorations

Environment Installation Using Anaconda

    conda create --f environment.yml

Visualization

Tensorboard logged datas will be located below runs/ directory, to visualize data after a pome run

    tensorboard --logdir runs or
    python -m tensorboard.main --logdir=./experiments/{algo}/runs

Todos

1. Currently, one file implementation pome.py can run without error, while algorithm is not tested
2. For reproduction of OpenAI Baselines, there are many addtional implementation details, according to this site
3. Visualization
4. code for other environments

Current Implementation

1. fixed-length trajectory segments
2. Orthogonal Initialization of Weights and Constant Initialization of biases
3. Mini-batch Updates
4. Skip Frame
5. Resize images
6. Scaling the Images to Range [0, 1]
7. minibatch standardization
8. remove reward estimation