Machine_ToM

The Implementation of "Machine Theory of Mind", ICML 2018, You can read the paper http://proceedings.mlr.press/v80/rabinowitz18a/rabinowitz18a.pdf

Contents

• 1. Install
• 2. Code Structure
• 3. Run the Code
• 4. Experiment Description

Install

• will be update soon

Structure

└─Machine_ToM
    ├─agent : Agent directory used in Experiments
    ├─environment : Environment
    ├─experiment :  Experiment files
    ├─model : Machine ToM models
    └─utils : dataloader, storage and visualization

Run the code

python main.py --num_exp 2 --sub_exp 1 --num_epoch 1000

Check the environment, agent, etc

• Environment

python environment/env.py

• Agent

python agent/reward_seeking_agent.py

Experiment Description

• Experiment 1: In this experiment, we predict the future action of current state with random agents whose policies are depending on Dirichlet dist. You can adjust the number of past trajectory by --sub_exp.
• Experiment 2: In this experiment, we predict the future action, consumption, successor representation of value iteration agents. The number of walls is sampled between 0 and 4.

There are three sub experiments.

• first sub experiment : MToM with the full trajectory of an agent on single past MDP. Agent gets a panalty(-0.01) for every move.
• second sub experiment : MToM with partial trajectory(one step) of an agent on single past MDP. Agent gets panalty(-0.01) for every move.
• third sub experiment : same as first sub experiment. But agent gets high panalty(-0.05) for every move.