Notes 8 27 2021 - opendigital/RL-collective-action GitHub Wiki

New states:

• Agent only has access to last round
• Agent has access to whole game

New reward functions:

• Sum Bush-Mosteller agents' contributions only
• Proportion of contributions by Bush-Mosteller agents that are > 0.5

Have 5 experiments:

• 2 reward functions x 2 state
• 2 baselines: 3 and 4 Bush-Mosteller agents

Results analysis:

• Compare total contributions of ALL 4 agents (including DQN agent) in experiments to 4-Bush-Mosteller baseline
• Compare Bush-Mosteller agent contributions in experiments to 3-Bush-Mosteller baseline
• Use simple t-test to prove that DQN agent optimizes contributions