Anna Deichler

The thesis treats situated communication as two coupled problems — producing spatially grounded nonverbal behavior, and interpreting it. It contributes generative models for co-speech and pointing gesture (CSMP, diffusion; imitation+RL), models and benchmarks for resolving referring expressions against a 3D scene (MM-Conv, PoseRefer), and the datasets underneath — controlled mocap and ego-exo capture with AR glasses.

AlphaGo was the first system to reach professional human-level play in Go by combining pattern knowledge from a deep neural network with Monte Carlo tree search (MCTS). MCTS relies on local, dynamic position evaluation — each edge stores its own statistics, making it easy to separate the effect of actions locally — whereas the strength of deep networks is generalization from previous experience to new situations. The pairing echoes the dual-process account of human reasoning: search plays the role of slow, deliberate planning, the network that of fast intuition, and the success of AlphaZero can be seen as the complementary interplay of the two.

The thesis examined the relative importance of local search versus generalization in AlphaZero in single-player, deterministic, fully observable RL environments (OpenAI Gym, PyBullet). The trade-off was probed by varying the number of MCTS iterations N_MCTS per decision step: under a fixed time budget this sets how much effort goes into accurate local values via larger search trees versus improving generalization through more frequent network updates.

Mid-size trees achieved the best performance for fixed N_MCTS. Adaptively choosing N_MCTS at each decision step — comparing root return variance against a rolling baseline and running extra iterations when value estimates are uncertain — improved both computational efficiency and final performance.

The thesis examined system stability under different hold functions in digital control, with and without time delays, for the classical problem of balancing an inverted pendulum with a discrete-time PD controller. Stability was analyzed for zero-order, first-order, second-order, and system-matched hold (SMH) functions — SMH being a special generalized sampled-data hold derived from the system dynamics — and presented as stability charts in the plane of proportional and derivative gains.

Higher-order hold functions enlarge the stable region in the gain plane, and with SMH the stable region becomes infinite when no time delay is present. Delays significantly shrink the stable region in all cases; the critical (smallest stabilizable) pendulum length was computed for zero-order hold, and higher-order holds were shown to decrease it for a given delay.