-
Notifications
You must be signed in to change notification settings - Fork 0
Reflections
Overall, a large variety of approaches were investigated in creating three separate agents. These approaches were described from a theoretical point of view, and the specific implementation within each agent were described, however the only approach that was not utilised within an agent was Approach A - PDDL / Classical Planning. This was mainly due to technical issues that were experienced in attempting to make calls to the solver (Metric-FF). Excluding this failed implementation, all agents that were built provided interesting Pacman gameplay, and allowed us to experiment with new ideas which would ultimately lead to our final agent: Agent 3 (A* Heuristic Search and Behaviour Trees).
If time permitted, the following AI methods would be of interest for further investigation (in addition to building upon and refining the current agents):
- Monte Carlo Tree Search or UCT;
- Deep Q-Learning;
- Goal Recognition techniques;
- Game Theoretic Methods.
This project provided us with an goal: to develop an autonomous Pacman agent team. In working towards this goal, we were able to plan; investigate; research, and implement many different AI methods, and in doing so, solidify the content taught throughout COMP90054 AI Planning for Autonomy. Furthermore, this project pushed us to collaborate in completing this goal as a team, building upon collaborate brainstorming; coding, and other general teamwork skills. It is, therefore, the joint opinion of team intelligence-artificial that this project was both informative and conducive to the learning outcomes outlined throughout this subject.