settings.py

'''All parameters are set here'''

class AgentSetting():
	
	''''commented values are the used ones in the nature paper, use them if you got that much of computing power'''

	minibatch = 32
	
	replay_strt_size = 15000#50000 # at start use randm policy for this time to fill memory b4 learning
	replay_memory = 50000#1000000
	
	t_net_update_freq = 10000 #10000
	
	discount_factor = 0.99
	
	update_freq = 1 #4 #update net every 4 actions
	
	#nature paper:50 million frames to be trained on 
	#double DQN paper: 200 million frames,ie, 50M steps 
	training_steps = 1000000
	
	#RMSProp 
	learning_rate = 0.00025 #not dueling
	duel_learining_rate = 0.0000625  # dueling lr (6.25*10^-5)

	momentum = 0.95
	#deprecated
	#grad_momentum = 0.95
	#sq_grad_momentum = 0.95
	#min_sq_grad = 0.01

	#epsilon
	e_greedy_init = 1.0
	e_greedy_final = 0.1

	e_final_at = 100000#1000000 #training_steps // 50 #1000000 #steps over which epsilon is annealed to its final value

	#TODO apply evalution every 1M steps!
	no_op_max = 30 #max of no-op action
	eval_every = 1000000 #steps
	epsilon_eval = 0.05

class ArchitectureSetting():
	
	#layer 1
	in_shape = [84,84,4]
	f1_no  = 32
	f1_size = [8,8]
	stride1 = 4
	#layer2
	f2_no = 64
	f2_size = [4,4]
	stride2 = 2
	#layer3
	f3_no = 64
	f3_size = [3,3]
	stride3 = 1
	#layer4 fc
	nodes = 512  #if dueling 512 for value and same for adv



class StateProcessorSetting():
	
	history_length = 4
	observation_dims = [84, 84]
		
class EnvSetting():
	
	recEvery = 100 #rec every ? episode
	'''gym:Each action is repeatedly performed for a duration of k frames, where k is uniformly sampled from {2, 3, 4}.'''
	action_repeat = 1 #4 
	
	#4 reward clipping
	max_reward =  1.0
	min_reward = -1.0

	#display
	frame_dim = [210,160,3]
	render = True


class UtilSettings():
	'''useful paths '''

	trainDir = 'graphVars/train' #dir for checkpoints during training
	playDir = 'graphVars/play' #dir for model weights for playing/eval
	monitorDir = 'gymRecordings'
	#experienceDir = 'expMemory' #TODO-kill
	trainSummaryDir = 'summaries/train'
	playSummaryDir = 'summaries/play'

class PerSettings():
	# doubleQ (tuned) , per-> proportional method
	#priority paras
	epsilon = 1.0#0.1 #0.01
	alpha = 0.6

	#importance sampling paras
	beta_init = 0.4
	beta_final = 1.0 #linear annealing till end of training
	beta_finalAt = AgentSetting.training_steps

	step_size = AgentSetting.learning_rate / 4.0