Doctoral Dissertation Defense

Scalable Planning through Learning in Fully and Partially Observable Domains

Rajesh Mangannavar
Advisor: Prasad Tadepalli
Committee: Alan Fern, Stefan Lee, Michael Rosulek
Oregon State University · April 28, 2026
"Small problems, when properly represented and learned from, carry enough structural information to guide decision-making at much larger scales."

Planning — choosing a sequence of actions to reach a goal — is fundamental to intelligent behavior. But it is computationally hard: search spaces explode exponentially with problem size, and partial observability makes it worse. This thesis introduces two complementary strategies that exploit planners' effectiveness at small scale.

Strategy 1

Learn from Small Problems

Train on small instances where planners produce optimal solutions. Transfer learned knowledge to larger instances where direct planning is too expensive.

Strategy 2

Abstract Large Problems

Use structured abstraction to reduce a large problem to a smaller one the planner can handle directly.

Contribution 1 — Fully Observable

GABAR: Graph-Based Action Ranking

NeurIPS 2025 (Main Conference)

Classical planners explore vast search trees to find plans. GABAR replaces this expensive search with a learned policy that directly ranks actions, generalizing from small training problems to instances 8× larger.

How It Works

Each planning state is converted into an action-centric graph where objects, predicates, and actions become nodes with typed edges. A GNN encoder processes this graph, and a GRU decoder sequentially ranks action schemas and their parameters — no search required.

State → Graph → GNN → GRU Decode → Action
89.2%
Coverage on hard problems (8× training size)
8/8
IPC benchmark domains
14×
Gap over value-based baselines
>LLMs
Outperforms state-of-the-art LLM planners

Action-centric graph representation is the most critical component — ablating it causes the largest performance drop

Action ranking with conditional decoding outperforms learning global value functions

100% coverage on Blocksworld, Gripper, and Miconic across all difficulty levels

Contribution 2 — Partially Observable

GammaZero: Belief-Space Search with Graph Representations

ICAPS 2026

When the world is only partially observable, agents must reason over beliefs (probability distributions over possible states) rather than states directly. GammaZero represents these beliefs as graphs, enabling a GNN to learn value functions and policies that generalize across problem sizes — something prior methods like BetaZero cannot do.

How It Works

Particle beliefs are aggregated into uncertainty-aware graphs where object attributes become node features weighted by belief probability. A GNN learns both a value function V(b) and a policy P(b) that guide Monte Carlo Tree Search, replacing expensive random rollouts with a single forward pass.

Belief → Graph → GNN → V(b), P(b) → MCTS
2–6×
Zero-shot generalization to larger problems
≥BetaZero
Matches SOTA on same-size tests
4
POMDP benchmark domains
20×
Computational efficiency vs. classical planners

Belief-to-graph construction encodes uncertainty directly in the graph topology

Size-invariant representation — the same trained model works on problems it has never seen

Extends GABAR's graph-based learning paradigm into the partially observable setting

Publications

NeurIPS 2025

Graph Neural Network Based Action Ranking for Planning

Rajesh Mangannavar, Stefan Lee, Alan Fern, Prasad Tadepalli

PDF Project

ICAPS 2026

GammaZero: Learning to Guide Belief-Space Search with Graph Representations

Rajesh Mangannavar, Prasad Tadepalli

Project

NeurIPS SpaLVE 2025

Hierarchical Object-Oriented POMDP Planning for Object Rearrangement

Rajesh Mangannavar, Alan Fern, Prasad Tadepalli

PDF Project

Thesis document available upon request

Interactive Defense Presentation

Walk through the full 44-slide defense with animated visualizations, step-through diagrams, and keyboard navigation — just like the live defense.

→ Launch Presentation