The next generation of AI systems should not only solve fixed tasks. They should be able to discover useful possibilities, adapt when conditions change, and remember what they have learned so that future learning becomes easier. This matters for autonomous agents and robots, but also for scientific and engineering problems where evaluations are expensive, uncertainty is high, and progress often depends on finding the right stepping stones.

In this talk, Dr. Vassiliades will present his research toward this goal. He will start from the view of discovery as adaptive search: generating diverse possibilities, evaluating them, retaining useful solutions, and reusing them later. He will show how this idea appears in quality-diversity methods, where the aim is not only to find one best solution but to build archives of diverse, high-performing behaviours or designs. Dr. Vassiliades will then discuss how prior knowledge, model-based learning, and Bayesian optimization can support efficient adaptation when real-world interaction is limited or costly. Finally, he will connect these ideas to lifelong learning, where the challenge is to retain, update, and reuse knowledge over time.

Dr. Vassiliades will conclude with a broader vision for AI systems that close the loop between search, adaptation, and memory: open-ended autocurricula in which agents and tasks co-adapt, hierarchical world models for lifelong skill learning, and adaptive discovery loops that help decide which simulations, experiments, designs, or hypotheses are worth exploring next.

Dr. Vassilis Vassiliades is a Research Assistant Professor at CYENS Centre of Excellence in Cyprus, where he leads the Cognitive Artificial Intelligence and Robotics research team. He received an MSc in Intelligent Systems Engineering from the University of Birmingham, UK (2008), and a PhD in Computer Science (Artificial Intelligence) from the University of Cyprus (2015). He completed postdoctoral research at Inria, France (2015-2018), where he worked on quality-diversity algorithms and data-efficient robot learning.

His research focuses on software agents and robots that efficiently learn from multimodal data, adapt to changing conditions, and accumulate skills over time in complex environments. His work has been supported by national, European, and industry-funded projects, and has appeared in international venues across machine learning, neural networks, evolutionary computation, robotics, and multiagent systems, with applications in interdisciplinary domains.