Saliency Aware Temporal Difference Learning

Motivation

The idea behind this project is to evaluate changes in the behaviour of an agent when mechanisms similar to incentive saliency are introduced inside a simple Temporal Difference Learning algorithm.

Agents equipped with incentive salience mechanisms will be prone to show a bias towards saliency-prone rewards (i.e., violet boxes) while they will treat other rewards (i.e., golden boxes) normally.

Features

• Create environments using text files.
• Parametrically vary the susceptibility of an agent to saliency effects.
• Simulate the behaviour of different agents within different environments.
• Analize the results and visualize the behaviour of agents.

Results

The Grid

Normal Agent

normal_grid_10000.mp4

Biased Agent

addicted_01_grid_10000.mp4

The Wall

Normal Agent

normal_wall_10000.mp4

Biased Agent

addicted_01_wall_10000.mp4

The Triple Wall

Normal Agent

normal_triple_wall_10000.mp4

Biased Agent

addicted_01_triple_wall_10000.mp4

The Treasure Island

Normal Agent

normal_treasure_island_10000.mp4

Biased Agent

addicted_01_treasure_island_10000.mp4

The Maze

Normal Agent

normal_maze_10000.mp4

Biased Agent

addicted_01_maze_10000.mp4

The TBone

Normal Agent

normal_tbone_10000.mp4

Biased Agent

addicted_01_tbone_10000.mp4