mavi

Main entry points:

Fairly nasty environment-class (visualization could really use a cleanup) in

• mazeenv/maze_environment.py

Example 'agent' which computes a randomized value function and visualize it:

• basic/viagent.py

Course exercises (very little of which is needed)

• irlc

Tasks + plan

pip install gym==0.21.0

• Foerst: Faa almindelig value-iteration til at fungere i Mazebase-gridworld. Dvs. vi antager at vi kender dynamikken p_ij og r_ij, r_ij^out, og planlaegger paa en horizont paa K. Dvs. vi vaelger phi-funktionen i VP paperet og implementerer algoritmen i numpy. Den skal kunne planlaegge at gaa til et maal optimalt.
• Implementer en paralell torch-version af koden og check at de to implementationer giver samme output. Dette check kan senere bruges til at lave en effektiv torch-version.
• Actor-critic (dvs. foerste algoritme i section 4). Check at vi kan navigere maze-environmentet.
• Multiagents: Speaker-listener opgaven (simpleste ikke-trivielle)
• Actor-critic baseline (samme implementation som actor critic men v(s) er et almindeligt 2-lags NN)