Even the creators of AI are often unable to read from text records and program statements exactly why the model made the decision it made. Today, this technology controls robots, autonomous vehicles, or entire industrial systems. The PANSim tool from CTU, which won a prestigious award from world experts at the ICAPS conference in Dublin, now helps to better understand the complex decision-making process.
The tool, which was created as a bachelor's thesis by Erol Medenčević, made the choices of artificial intelligence clearer in an interactive 2D simulation. To make the presentation as understandable as possible, the authors chose a simple example: an AI agent must, in the role of an animated frog, get across a pond and collect all the coins along the way, while some water lilies may disappear below the surface at any moment. When solving this task, developers can watch how their agent copes with random changes in the environment and what further steps it plans to take to achieve the goal. “It’s not just about the graphics themselves. The visual appearance allows a much better understanding of what is actually happening during the decision-making process. Thanks to this, the developer quickly recognizes why the agent behaved in a certain way and where the algorithm needs to be adjusted,” said the supervisor of the bachelor’s thesis, Jakub Med.
The tool’s potential was fully demonstrated at this year’s conference in Dublin, which brought together the elite of experts in automated planning. Out of 12 projects from leading world universities and research institutions, PANSim received the most votes. “When my colleagues sent me a photo of the diploma for the best demonstration from Dublin, I was really surprised. It’s a great motivation to continue,” said Erol Medenčević, who is continuing his master’s degree in the development of PANSim.
The authors now want to make AI decision-making processes even more transparent, add the ability to go back a few steps in the simulation or expand the tool with additional scenarios. The long-term goal is to create a universal environment that will allow for the visualization of a wide range of planning tasks.