The DT8124 PhD-level course on Intelligent Agents introduces foundational and advanced concepts in agent-based reasoning, planning, and motion. It addresses doctoral candidates from Computer Science, Engineeringp Cybernetics, Robotics, Mechanical Engineering, and Maritime Technology—fields that increasingly rely on autonomous systems capable of perceiving, reasoning, and acting robustly in complex environments.
The course begins with the conceptual and theoretical foundations of agents: what it means to behave intelligently, how rationality is formally defined, and how tasks, environments, and goals can be modeled using utility-based and logical approaches. These fundamental ideas are embedded in a broader understanding of how intelligent behavior arises from an agent’s architecture—its internal structure for perception, decision-making, and action.
Several architectural paradigms are presented and contrasted, including:
The course also covers formal reasoning methods, including symbolic deduction, practical reasoning, and planning with logical operators. Participants engage with well-established formalisms such as STRIPS and BDI (Belief-Desire-Intention) models to understand how goal-directed behavior can be systematically structured.
A major emphasis is placed on planning and motion, addressing both discrete and continuous decision-making. Techniques covered range from symbolic planners and graph-based algorithms (like A*) to sampling-based motion planners (e.g., PRMs, RRTs), and dynamic re-planning under uncertainty. Additionally, reactive approaches like the Dynamic Window Approach (DWA) are studied in the context of real-time control under dynamic constraints.
To bridge toward modern applications, we also provide an overview of selected machine learning approaches relevant to agent control, including attention-based architectures and diffusion-inspired techniques, highlighting their growing influence in planning and motion synthesis.
Overall, the course provides a rigorous and comprehensive foundation for understanding and designing intelligent agents across diverse physical and virtual domains.
Basic definitions of agents and environments; sensing, acting, autonomy. Rationality and performance measures. Agent architectures as combinations of perception, decision-making, and actuation. Agent types: purely reactive, state-based, utility-based.
Agent-environment interaction as state-action histories; transition models and agent functions. Utility-based evaluation of agent performance. Specification of achievement and maintenance tasks via predicates and utility functions. Notions of bounded optimality and agent synthesis.
Design paradigms from purely reactive to hybrid deliberative/reactive architectures. Symbolic agents using deductive reasoning. Practical reasoning agents based on Beliefs, Desires, and Intentions (BDI). Reasoning as search: deduction, abduction, induction, and analogical inference.
Languages for agent specification and execution: Agent0, PLACA, Concurrent MetateM. Mental state representations and rule-based behavior. Temporal logic constructs and execution models. Foundations for interpretable and verifiable agent behavior.
Planning as means-ends reasoning in intentional agents. Planning languages and formalisms, especially STRIPS. Partial-order planning, plan representation, causal links, clobbering and promotion. Planning in symbolic domains like the blocks world.
State-space search methods for planning. Heuristic guidance using relaxed problems. Planning graphs and the Graphplan algorithm. Efficient encoding of planning constraints and mutex relationships.
Reactive behavior from subsumption architectures to behavior trees. Hybrid architectures combining planning and reactivity (e.g., Touring Machines, 3T). Modern implementations in robotic middleware (e.g., ROS-based systems, behavior-based controllers, MoveIt! and Navigation Stack).
Configuration space concepts for mobile and articulated robots. Combinatorial planning methods (visibility graphs, Voronoi, cell decomposition). Sampling-based methods: PRM, RRT, and their variants. Recent extensions involving adaptive sampling and learning-based planning policies.
Approaches for planning in changing or uncertain environments. Any-angle A*, D* and D* Lite for efficient replanning. Integration of kinematic and dynamic constraints via 5D planning and trajectory optimization. Emphasis on real-world planning challenges.
Dynamic Window Approach (DWA) and motion selection in (v, ω)-space. Predictive control under dynamic constraints. Fusion of model-based control (e.g., MPC) with learned motion policies for short-term planning. New approaches that combine simulation, learned value functions, and control-theoretic insight.
Fundamentals of decision-theoretic agent design based on Markov Decision Processes (MDPs); state utilities, policies, and the Bellman equation; value iteration and dynamic programming for optimal control.
Applications to uncertain action outcomes, stochastic dynamics, and incomplete knowledge of the environment. Examples include robust behavior in mobile robots, policy execution under noise, and handling action failure.
The course highlights the role of policy-based approaches as robust alternatives to brittle action sequences, and introduces the use of simulation-based learning, deep value estimation, and policy approximation.
While classical formulations are emphasized for clarity and theoretical grounding, we also explore recent advances such as deep reinforcement learning and world model-based planning, showing how uncertainty modeling can be fused with neural policy learning and predictive simulation.
Short outlook on transformer-based architectures and diffusion models for sequential decision-making and motion generation. While not central to the course, these methods are briefly discussed in the context of their growing relevance for planning and policy learning in intelligent agents.
The DT8124 course is organized as a weekly, three-hour in-person event held on campus, typically starting in late August and running through the end of November. The course consists of approximately 13–14 sessions. Participants may miss up to 10% of the weekly meetings without affecting course approval (1 event, or 2 on explicit prior agreement).
Each session generally includes a 90-minute lecture followed by an instructor-led discussion or exercise. These post-lecture sessions may involve collaborative exercises, homework discussions, or joint attendance at live guest lectures. Guest speakers are invited to present their work and engage in discussions with the participants, either in person or via live remote connection.
In the first phase of the course, each participant selects a personal topic—ideally connected to both the course contents and their own PhD project. Weekly homeworks accompany the sessions and may consist of extended reading, short writing assignments (1–2 pages), or both. These tasks are tailored to the participants' observed interests and group discussions, providing a dynamic and responsive learning structure.
Participants should expect to dedicate 5-6 hours per week in addition to class attendance to reading, homework, and preparation. Additional to that, comes the time for preparing the individual presentation and the term paper plus exam preparation. Specific formatting and submission guidelines for assignments and the term paper will be provided during the course.
Individual presentations per participants, which are a mandatory and graded component of the course, begin in late September or early October and are scheduled based on the number of participants. Presentations focus on surveying a relevant research area, and presenters receive real-time feedback from both instructors and peers during class discussions.
The term paper may expand upon the presentation topic, but should represent a more in-depth and original contribution. For a survey-style paper, a longer format (approximately 12 pages of net content plus references) is expected. Otherwise, participants submit a 6-page paper in the style of a research conference submission. Feedback on term papers is given individually to the authors.
The course concludes with an oral exam, usually held in December. The exam lasts approximately 30 minutes (extended to 40 minutes if needed). Further details regarding grading and formal examination procedures can be found in the official NTNU course description.