1-Dependencies: Numpy Scipy Matplotlib Pandas Math Random Tensorflow Keras Pytorch

2-Results:

a)Sea level rise projections:

Regionally adjusted sea level rise projections from the National Oceanic and Atmospheric Administration (NOAA) is available in the link: http://www.tbrpc.org/wp-content/uploads/2019/05/CSAP_SLR_Recommendation_2019.pdf We obtained our fitted curve with help of the table in page 6 of the above pdf. After converting the feet values to millimeter, we extracted the annual relative sea level. This is implemented in NOAA_fittings.ipynb

c)MDP Based Agent:

The government cooperation index a_g, the residents’ cooperation index a_r, and the sea level projection selector a are user defined inputs. The user inputs a=0,1,2, respectively for intermediate low, intermediate, and high sea level rise projections. We use the standard DQN algorithm with experience replay. This is implemented using DQN in slr_dqn.ipynb and Advantage Actro Critic algorithm in slr_A2C.ipynb

(i)DQN Hyperparameters: GAMMA = [0.1 , 0.3, 0.5, 0.7, 0.9] *** LEARNING_RATE = 0.001 MEMORY_SIZE = 20000 BATCH_SIZE = 1000 EXPLORATION_MAX = 1.0 EXPLORATION_MIN = 0.01 EXPLORATION_DECAY = 0.9999 *** Gamma corresponds to the government cooperation index a_g, so we used it as an input to the function.

DQN Model structure: Dense layer - input: 2, output: 12, activation: relu Dense layer - input: 12, output: 24, activation: relu Dense layer – input: 24, output: 12, activation: relu Dense layer – input: 12, output: 4, activation: linear Huber loss function Adam optimizer State (ln,sn) is a (1,2) size array Action is integer between 0 and 3 Cost has a positive continuous value Next State (ln+1,sn+1) is a (1,2) size array Every 50 steps target network's weights are updated

(ii)Advantage Actor Critic(A2c) Hyperparameters: GAMMA = [0.1 , 0.3, 0.5, 0.7, 0.9] ***

*** Gamma corresponds to the government cooperation index a_g, so we used it as an input to the function.

Actor Model structure: Dense layer - input: 2, output: 12, activation: relu Dense layer - input: 12, output: 120, activation: relu Dense layer – input: 120, output: 12, activation: relu Dense layer – input: 12, output: 4, activation: softmax LEARNING_RATE = 3e-4r

CriticModel structure: Dense layer - input: 2, output: 12, activation: relu Dense layer - input: 12, output: 120, activation: relu Dense layer – input: 120, output: 12, activation: relu Dense layer – input: 12, output: 1, activation: linear LEARNING_RATE = 3e-4r

State (ln,sn) is a (1,2) size array
Action is integer between 0 and 3
Cost has a positive continuous value
Next State (ln+1,sn+1) is a (1,2) size array