Reinforcement Learning in Third-Person Environments: A Hybrid Q-Learning and Sensorimotor Approach

Q: How many days will it take for my paper to be published?

The review time for papers is not fixed. However, if the paper is accepted and the author completes the processing charges formalities, the paper will be published within a few working days.

Q: I would like to receive a hard copy of the journal materials. Are there any additional charges?

You can log in to the author portal and pay 500 INR to receive the hard copy materials.

Harsh Mittal; Kumar Namah; Ashmit Pandey

Reinforcement Learning in Third-Person Environments: A Hybrid Q-Learning and Sensorimotor Approach

Authors: Harsh Mittal, Kumar Namah, Ashmit Pandey

Unique Paper ID: 179239
Volume: 11
Issue: 12
PageNo: 8136-8140

Keywords: Reinforcement Learning Q-Learning Sensorimotor Control Devil May Cry 3 Game AI Style Evaluation Vision-Based Learning Modular Training

Abstract:
The area of reinforcement learning (RL) has evolved significantly, particularly in first-person games and simulation environments like VizDoom, in which agents can be trained directly from raw pixels. Yet transferring RL to third-person action games, like Devil May Cry 3, presents additional challenges such as partial observability, high-dimensional pixel data, and the need for stylistic fighting behavior. This work introduces a learning framework that, in contrast to earlier methods dependent on raw sensor feedback, uses learned latent dynamics (world models) to model game worlds and train policies in a dense, imagination-based representation space. A convolutional neural network processes gameplay frames to obtain low- dimensional latent vectors describing key environmental features, including opponent positions and player positions. To predict future states and rewards, the hidden representations are then passed to a Recurrent State-Space Model (RSSM). In this hidden space, an actor-critic reinforcement learning agent decides the optimal action by finding a compromise between short-term tactics and long-term planning. As compared to end-to-end pixel- based learning, this process vastly improves training efficiency while allowing adaptive play. The reward system is optimized not only for success, but also for flair—obtaining high-ranked performances (e.g., S-rank) by means of combo diversity and spatial awareness. A modular training paradigm is employed, partitioning combat, exploration, and boss battles into separate learning tasks, enabling targeted optimization and later integration. This method shows enhanced agent performance, flexibility, and potential generalization to other visually rich, multi-objective tasks. The results are used to advance AI technology towards real- time decision making in visually dense settings with changing objectives.

Download article

email to a friend

Cite This Article

ISSN: 2349-6002
Volume: 11
Issue: 12
PageNo: 8136-8140

Reinforcement Learning in Third-Person Environments: A Hybrid Q-Learning and Sensorimotor Approach

Available:https://ijirt.org/article?manuscript=179239

Impact Factor
8.01 (Year 2024)

UGC Approved
Journal no 47859

Join Our IPN

IJIRT Partner Network

Submit your research paper and those of your network (friends, colleagues, or peers) through your IPN account, and receive 800 INR for each paper that gets published.

Join Now