Add config file to algorithm scc

examples/ma_cases/scc.yaml

@@ -0,0 +1,85 @@
+alg_para:
+ alg_name: SCCAlg
+ alg_config: # algorithm config
+ batch_size: 32 # train batch size
+ buffer_size: 1000 # train buffer size of trajectory
+ epsilon_anneal_time: 50000 # time_length of epsilon schedule
+ epsilon_finish: 0.00 # epsilon minimum
+ epsilon_start: 1.0 # epsilon maximum
+ obs_agent_id: True # use agent id within state
+ obs_last_action: True # use last action within state with onehot
+ target_update_interval: 200 # interval to update target network
+
+env_para:
+ env_name: StarCraft2Xt
+ env_info: {
+ "continuing_episode": False,
+ "difficulty": "7",
+ "game_version": null, # "latest",
+ "map_name": "2s_vs_1sc",
+ "move_amount": 2,
+ "obs_all_health": True,
+ "obs_instead_of_state": False,
+ "obs_last_action": False,
+ "obs_own_health": True,
+ "obs_pathing_grid": False,
+ "obs_terrain_height": False,
+ "obs_timestep_number": False,
+ "reward_death_value": 10,
+ "reward_defeat": 0,
+ "reward_negative_scale": 0.5,
+ "reward_only_positive": True,
+ "reward_scale": True,
+ "reward_scale_rate": 20,
+ "reward_sparse": False,
+ "reward_win": 200,
+ "replay_dir": "",
+ "replay_prefix": "",
+ "state_last_action": True,
+ "state_timestep_number": False,
+ "step_mul": 8,
+ "seed": 22,
+ "heuristic_ai": False,
+ "heuristic_rest": False,
+ "debug": False,
+ }
+
+agent_para:
+ agent_name: StarCraftSCC
+ agent_num : 1
+ agent_config: {
+ 'complete_step': 2050000, 
+ }
+
+model_para:
+ actor:
+ model_name: SCCModel # The qmix model defined within xingtian
+ use_npu: False # npu usage flag
+ allow_mix_precision: True # setup mix precision
+
+ model_config:
+ gamma: 0.99 # discount value for accumulative reward
+ mixer_grad_norm_clip: 5 # mixer value for grad norm
+ actor_grad_norm_clip: 5 # actor clip value for grad norm
+ a_lr: 0.0005 # actor learning rate
+ c_lr: 0.0005 # mixer learning rate
+ rnn_hidden_dim: 64 # the dimensionality of rnn hidden
+ batch_size: 32 # train batch size for tensorboard model build
+ use_double_q: True # build model with double q algorithm
+ dense_unit_number: 128
+ enable_critic_multi_channel: True
+ channel_merge: 'concat' # concat or add
+ mc_sample_times: 3
+ map_name: "2s_vs_1sc"
+
+
+env_num: 1 # explore environment number to parallel
+benchmark:
+ id: xt_scc
+ # archive_root: scc_results # default: ~/xt_archive
+ eval:
+ gap: 256 # train times call once evaluate
+ evaluator_num: 1 # run eval with how much evaluator instance
+ episodes_per_eval: 32 # run how much episodes within one evaluate
+ max_step_per_episode: 128
+