td3_implementation analysis

[CUN-bjy]

the first TD3 implementation do not work well..

so.. have to analysis each part of the differences from ddpg

[CUN-bjy]

[TEST 1]

• rebase to ddpg
• add target policy smoothing term.
• test on RoboschoolInvertedPendulum-v1.
• batch_size -> 64, hidden_layers -> 24, 16(for each actor and critic)
• lr-> 1e-4,1e-3, tau->1e-3,1e-3(for each actor and critic)
• it works.(cause of exploration problem, I tested 5 times more.. success rate is under 30%.)
  ✔️

[TEST 2]

• basically, tested on InvertedPendulum, almost same to above parameters and added target policy smoothing.
• change parameters, lr->3e-4,3e-4, tau->5e-3,5e-3 (same to original td3 code)
• it works.(also have exploration problem, success rate is not high.)
  ✔️

[TEST 3]

• add delayed policy update term.
• delayed update peoriod -> 2
• it works.(have exploration problem, success rate is too low, and feel like delayed to learn a good policy)
  ✔️

[CUN-bjy]

[TEST 4]

• rebase to ddpg
• add target policy smoothing term.
• test on RoboschoolInvertedPendulum-v1.
• batch_size -> 64, hidden_layers -> 24, 16(for each actor and critic)
• lr->3e-4,3e-4, tau->5e-3,5e-3 (for each actor and critic)
• (reset policy update interval to 1)
• add double cliped Q update term
• but only use Q1 for target update
• it works.
  ✔️

[TEST 5]

• same to above experiment.
• use Q1,Q2 for target update
• it doesn't work well..

[TEST 6]

• test on RoboschoolInvertedPendulum-v1.
• TD3 set.
• add target policy smoothing term.
• add delayed policy update term. update_interval -> 2
• add double cliped Q update term
• batch_size -> 64, hidden_layers -> 24, 16(for each actor and critic)
• lr->3e-4,3e-4, tau->5e-3,5e-3 (for each actor and critic)
• also doesn't work well..(take long time to learn the policy and getting worse)

👀 I think the my implementation of double cliped q update has some problems.

[CUN-bjy]

[TEST 7]

• test on RoboschoolInvertedPendulum-v1.
• add double cliped Q update term (only)
• batch_size -> 64, hidden_layers -> 24, 16(for each actor and critic)
• lr->3e-4,3e-4, tau->5e-3,5e-3 (for each actor and critic)
• yes.. doesn't work.. Catastrophic forgetting problem happen, even in one task.

[TEST 8]

• same condition to above.
• changed some codes that make independent optimizer for critics
  
  
• I think that is working, but very slow compared to simple ddpg.
• by double cliped q update overestimation problem is fixed, but that make policy update so slow.
• should try integrated system and than should change some parameters..

[CUN-bjy]

[TEST 9]

• integrate that all.. -> doesn't work..
  
  

[CUN-bjy]

[TEST 10]

• initial random policy added for exploration
• and there's some mistake..

before

		a = agent.make_action(obs,t)
		action = np.argmax(a) if is_discrete else a

		# do step on gym at t-step
		new_obs, reward, done, info = env.step(action) 

		# store the results to buffer	
		agent.memorize(obs, a, reward, done, new_obs) 
                # should've memorize action w/ noise!!

after

		a = agent.make_action(obs,t)
		action = np.argmax(a) if is_discrete else a

		# do step on gym at t-step
		new_obs, reward, done, info = env.step(action) 

		# store the results to buffer	
		agent.memorize(obs, action, reward, done, new_obs)

but, consequently, doesn't work..

[CUN-bjy]

[TEST 11]

• use OU noise process for off-policy exploration strategy
  
  

also, doesn't work.

[CUN-bjy]

[TEST 12]

• use OU noise process for off-policy exploration strategy
• without BatchNormalization, Weight Regularization, Groot Initializer on actor and critic.