In this project, an alternative architecture is proposed for the Fighting ICE competition which combines two of the most popular previous architectures reflex and learning based as subsystems. In this project I present a hybrid based architecture which combines two subsystems one reactive and one proactive and combines them into a single agent. A simple reflex agent is used to build in some expert knowledge and k nearest neighbor is used to predict attacks based on location. The sub-system is modeled as a vertically layered one pass architecture. This allows for the combination of different agent architectures at each layer and to choose the optimal action based on the perception system.
Follow the instructions on the Getting Started page which will shows you how to install Fighting ICE and the AIToolkit. You can then clone the repository and add the .jar file to the Fighting ICE project to verify the results and continue delveoping the source code.
The controller in it's current form is quite successful and has been able to beat a number of the controllers from the previous year. Further information about the design of the system can be seen in the report. The controller has only been tested on Version 2.0 of Fighting ICE and I have no plans to update to Version 3.0.
| Agent | Round 1 | Round 2 | Round 3 | Average |
|---|---|---|---|---|
| T3C | 418 | 757 | 726 | 634 |
| ATTeam2 | 478 | 415 | 518 | 470 |
| DragonKing3C | 447 | 493 | 578 | 506 |
This table shows the results against the top 3 controllers from last year with a score above 500 indicating a win.