rcln / dospordos

renforcement learning with Deep Q for highly qualified migrations localisations in Google-Bing-Duckduckgo

dospordos

Reinforcement learning using the technique of TD-Gammon, for highly qualified migrations localizations in Google-Bing-Duckduckgo-Citeseerx

How to run

The script is called training_script.py which has the following arguments:

• "DB", "Path to training directory"
• "ALG", "Algorithm to execute", default="DQN"
• "is_RE", "Use of Regular Expression", default="0"
• "-is_test", "The data is for testing", required=False, default=0
• "-initial_range", "Initial range of users", required=False
• "-final_range", help="Final range of users", required=False
• "-is_db_v2", help="Is the second database", required=False

Running Examples

python3 training_script.py ~/project/dospordos/DATA/db_v1_ns/train_db/ DQN 0

This means that you'll be using:

• the train database (data source) with path ~/project/dospordos/DATA/db_v1_ns/train_db/
• DQN model instead of a DDQN
• 0 is for using Regex or NE and is_test is for using a train data source or test data source.

python3 training_script.py ~/project/dospordos/DATA/db_v2_ns/test_db/ DQN 0 -is_test=1 -is_db_v2=1

This means that you'll be using:

• the test database db_v2_ns (data source)
• is the second database

Depending on the parameters given the data will be stored as DQN_0_db_v1_ns* in the DATA directory

There are other optional parameters to run a specific range of users which are -initial_range and -final_range

python training_script.py /users/urbinagonzalez/project/dospordos/DATA/db_v1_ns/test_db/ DQN 0 -is_test=1 -final_range=45

• This will run the users up to the user 45. Is doing list_users[:45]

Requirements

The data directory should have folders with numbers

• ~/DATA/train
  • 3
  • 5
  • ...
  • 4904
  • 4905
  • ...

Besides running the build.sh

• python -m spacy download en
• Install keras, tensorflow

####Notes

In the class DQN of DQN_implementation.py you can set the callbacks used for stopping the network.

self.callbacks = [agent.EarlyStopByLossVal(value=0.1), agent.EarlyStopping(patience=10)]

##For testing Use TESTS/evaluate_test_run script, you should already have all the pkl files you want to average and graph. This is the format you should follow:

python3 evaluate_test_run.py -r DQN_0_db_v1_ns_rm.pkl -acc DQN_0_db_v1_ns_acc.pkl -g 1

For more details:

python3 evaluate_test_run.py -h

Connection to cluster

ssh -p 60022 [email protected]

in GPU2 The directory for the project is

~/project/dospordos

The virtual environment is called venv-dospordos

There's a tmux session ready. To connect

tmux a -t base