"Server.log_every_n": 10,
"PoloWrappedReplayBuffer.batch_size": 32,
"PoloOutOfGraphReplayBuffer.solved_unsolved_ratio": 0.5,
"training_steps": 200000,
"train_every_num_steps": 10,
"game_buffer_size": 10,
"run_eval_worker": False,
}
params_grid = {
"get_env_creator.N": [40], # Here set problem size
"MCTS.episode_max_steps": [
51,
], # should be no less than N for ChainEnvironment
"EnsembleValueAccumulatorMeanStdMaxUCB.kappa_fn": [
ConstKappa(0),
],
"ValueBase.model_name": [
"linear_multi_head",
],
}
experiments_list = create_experiments_helper(
experiment_name='Ensemble ChainEnv',
python_path='.:./deps/gym-sokoban:./deps/ourlib:'
'./deps/baselines:./deps/dopamine:./deps/gym-sokoban-fast:./deps/chainenv:',
paths_to_dump='',
exclude=[],
base_config=base_config,
params_grid=params_grid)
"MCTS.num_sampling_moves": 0,
"MCTS.value_annealing": 1.0,
"MCTS.avoid_loops": True,
"MCTS.gamma": 0.99,
"MCTS.node_value_mode": "bootstrap",
"MCTS.episode_max_steps": 50,
"Server.min_replay_history": 1000,
"PoloWrappedReplayBuffer.batch_size": 32,
"PoloOutOfGraphReplayBuffer.solved_unsolved_ratio": 0.5,
"curriculum": False,
"training_steps": 500000,
"train_every_num_steps": 100,
"game_buffer_size": 25,
"log_every_n": 50,
"run_eval_worker": False,
"Server.save_checkpoint_every_train_steps": 500,
}
params_grid = {
"MCTS.avoid_loops": [True, False],
"ValueBase.model_name": ["convnet_mnist", "kc_parametrized_cnn_v0_2"]
}
experiments_list = create_experiments_helper(
experiment_name='Mcts sanity experiment',
python_path='.:./deps/gym-sokoban:./deps/ourlib:'
'./deps/baselines:./deps/dopamine:./deps/gym-sokoban-fast',
paths_to_dump='',
base_config=base_config,
params_grid=params_grid)
],
"KC_MCTS.num_ensembles_per_game": [
1,
],
"ValueBase.learning_rate_fn": [
0.00025,
],
"EnsembleValueAccumulatorMeanStdMaxUCB.ucb_coeff": [
0.0,
],
"EnsembleValueAccumulatorMeanStdMaxUCB.exploration_target": [
False,
],
"MCTS.num_mcts_passes": [
10,
],
"ValueBase.model_name": [
"multiple_mlps",
],
"ValueEnsemble2.prior_scale": [None],
}
experiments_list = create_experiments_helper(
experiment_name='Sokoban single-board',
python_path='.:./deps/gym-sokoban:./deps/ourlib:'
'./deps/baselines:./deps/dopamine:./deps/gym-sokoban-fast:./deps/chainenv:./deps/toy-mr:',
paths_to_dump='',
callbacks=(),
base_config=base_config,
params_grid=params_grid)
"EnsembleValueTraits.dead_end_value": -2.0,
"MCTSWithVotingTwoModels.num_mcts_passes": 10,
"MCTSWithVotingTwoModels.num_sampling_moves": 0,
"MCTSWithVotingTwoModels.avoid_loops": True,
"MCTSWithVotingTwoModels.gamma": 0.99,
"MCTSWithVotingTwoModels.episode_max_steps": 200,
"MCTSWithVotingTwoModels.avoid_history_coeff": -2.,
"Server.min_replay_history": 1000,
"PoloOutOfGraphReplayBuffer.solved_unsolved_ratio": 0.5,
"training_steps": 500000,
"train_every_num_steps": 100,
"game_buffer_size": 25,
"run_eval_worker": False,
"Server.log_every_n": 50,
"MCTSWithVotingTwoModels.node_value_mode": "bootstrap",
}
params_grid = {
"use_perfect_env.value": [False,],
"SimulatedSokobanEnvModel.model_path": ["checkpoints/epoch.0003.hdf5"],
"PoloWrappedReplayBuffer.batch_size": [96,],
}
experiments_list = create_experiments_helper(experiment_name='sokoban with learned model',
python_path='.:./deps/gym-sokoban:./deps/ourlib:'
'./deps/baselines:./deps/dopamine:./deps/gym-sokoban-fast:./deps/chainenv:./polo_plus/kc:./deps/toy-mr:',
paths_to_dump='',
base_config=base_config, params_grid=params_grid)