Deep Reinforcement learning benchmarks

Comparison of theano and tensorflow for deep RL.

Main idea is test arcitectures suitable for deep RL purposes i.e. not so deep CNN/FC networks optionally with LSTM layer.

Operations to compare:

• forward pass with batch size one for communicating with environment
• forward + backward pass with various batch sized (for training)

Test script runs all networks on GPU as well as CPU (1 thread and all available threads).

Architectures to compare:

1. FC network
2. DQN CNN aritecture
3. openai-universe-starter(openai) agent CNN arcitecture
4. LSTM network
5. LSTM with DQN convolution
6. LSTM with openai convolution

Running benchmark

To run benchmarks just use script:

$ ./run_all.sh

Optionally a separate experiment can be run with command

$ python run_benchmark.py

Use flag --help to see options for experiment.

Results

Raw results are in file results.txt. Preprocessed result can be found here. All results were obtained with GeForce GTX 980 and Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz.