Testing the Invariance of Skills and Strategies Developed by Artificial Agents under Different Sensory Modalities
No Thumbnail Available
Date
2024-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Artificial intelligence (AI) is the intelligence of machines or software, as opposed to the
intelligence of humans or animals by developing artificial intelligence agents. Artificial
Intelligence (AI) agent must be aware of the external environment to understand or to execute
tasks. The interaction between an agent and the environment is achieved by the aid of input sensory
modality (image, sound, lidar or other input modality) using metrics reward sizes they collect
during reinforcement learning. To do so, the we used two algorithms, Proximal Policy
Optimization (PPO) and Soft Actor-Critic (SAC) for the experiments. PPO is a reinforcement
learning algorithm that is used to train agents to perform tasks in environments through trial and
error. It is designed to optimize policies, which are the strategies or behaviors that the agents use
to make decisions. PPO aims to find the optimal policy by iteratively updating the parameters
based on the collected experiences from interactions with the environment. SAC is a reinforcement
learning algorithm used for training agents in environments with continuous action spaces. It is
an extension of the actor-critic framework that combines the advantages of both policy
optimization and value estimation methods. During the experiment researcher developed the
agents and the environment model in the project and selected the appropriate RL method,
implemented and designed the metrics of the learning performance of the agents depending on
their nature. We demonstrate agents' varied training performances under different sensory
modalities, with optimal outcomes observed when combining multiple modalities. Despite these
differences, the study underscores agents' adaptability across sensory inputs, advancing our
understanding of cross-modal learning in AI. Also, the researcher has shown the invariance of
sensory modality results on the learning skills and strategies of an agent and developed simple
game like reaching the goal with different input sensory information to test the agent skill and
strategies under different perceptual modalities.
Description
Keywords
Reinforcement Learning, Proximal Policy Optimization, Soft Actor-Critic