**Hi! I'm Muhammad Zaid, a Computer Science student
at FAST University majoring in Artificial Intelligence.
This is my technical blog post written as part of my
AI course. I'm passionate about how classical AI
algorithms connect to cutting-edge autonomous systems.
(@raqeeb_26 Looking forward to your best feedback !
**
Introduction
What happens when a 50-year-old search algorithm learns to think like an intelligent agent? And what does it look like when AI systems stop waiting for instructions and start pursuing goals autonomously? Two recent research papers answer these questions from complementary perspectives. Wang, Li, and Bain's "Research on the A* Algorithm Based on Adaptive Weights and Heuristic Reward Values" (2025) reimagines A* with dynamic heuristic weighting and reward-based escape mechanisms for autonomous vehicle path planning. Banda, Kangri, Nauru, Pastor, and Vivipara's "The Rise of Agentic AI" (2025) surveys 143 studies to map the emerging world of systems that plan, reason, reflect, and act autonomously. Together, they reveal a powerful progression: the intelligence we engineer into individual algorithms is the seed of autonomy defining next-generation AI.
Paper A : Making A* Adaptive and Intelligent
The Problem and Goal
The traditional A* evaluation function f(n) = g(n) + h(n) guarantees optimal paths with admissible heuristics but suffers in complex environments from excessive node expansion, slow computation, jagged paths, and unsafe corner-clipping near obstacles. Wang et al. solve all four problems through five interconnected innovations.
Technical Contributions with Course Connections
Coordinate-Based Five-Way Search with Safety Filtering
By computing Δx = x_ current − x_ goal and Δy = y_ current − y_ goal, the authors select only five goal-directed neighbors instead of eight, then filter diagonal moves near obstacles entirely.
Course Connection: This directly modifies A*'s successor generation step, reducing the branching factor from 8 to 5. Since A*'s time complexity is O(b^ d), this reduction produces exponential savings as depth increases. This is informed pruning — using domain knowledge to eliminate unpromising nodes before they enter the open list, a technique we discussed but never saw applied this elegantly. The obstacle-aware diagonal filtering transforms efficiency optimization into safety-aware search, something standard A* completely lacks.
Result: 10.5% fewer nodes searched, zero hazardous paths.Adaptive Dynamic Weighting via Radial Basis Functions
This is the paper's most brilliant contribution. Instead of a fixed heuristic weight, the authors introduce a spatially varying weight using a 2D Gaussian RBF:
f(n) = g(n) + w(x,y) · h(n)
where w(x,y) = A·exp(−[(x−μx)²/(2σx²) + (y−μy)²/(2σy²)]) + k
Near the start (center of the RBF), w ≈ 21 — the algorithm behaves like Greedy Best-First Search, rushing toward the goal. As distance from the start increases, w decays smoothly toward 1.0, and the algorithm transitions to standard A* behavior, exploring more carefully near the goal.
Course Connection: This profoundly bridges concepts we studied separately. In class, we learned Greedy Best-First Search (fast, suboptimal) and A* (optimal, slow) as discrete choices, with Weighted A* (f = g + ε·h) as a fixed compromise. The RBF creates a continuous, distance-dependent spectrum between these extremes — essentially an ε that decays spatially. The algorithm "knows" when to be aggressive and when to be careful, exhibiting contextual self-awareness that feels more like agent behavior than a static procedure. This directly addresses the admissibility vs. efficiency tradeoff we discussed, resolving it dynamically rather than statically.
Result: ~40% node reduction with only 2.2% path length increase.Heuristic Reward Value for Local Optima Escape
When long obstacles trap the search, all forward nodes may be closed. The authors pre-compute a reward field diffusing from the goal via BFS, adding it to the cost function:
f(n) = g(n) + w(x, y)·h(n) + t· reward_ value
The gradient guides search around obstacles by favoring cells with lower reward values.
Course Connection: This connects simultaneously to Artificial Potential Fields (attractive field from goal, implicit repulsion from obstacles), reward shaping in Reinforcement Learning (supplementary guidance beyond the standard heuristic), and heuristic relaxation (the reward values encode obstacle-aware distance — a second heuristic complementing Euclidean distance). Standard A*'s h(n) estimates straight-line distance but knows nothing about obstacle topology; the reward field fills this gap.
Result: Combined improvements achieve 76.2% node reduction versus traditional A*.
4–5. Path Post-Processing and Bezier Smoothing
Two-pass redundant node removal using vector cross-product intersection testing, followed by third-order Bezier curve smoothing, produces physically drivable paths. This mirrors iterative self-improvement cycles in agent learning.
Comprehensive Results
Validated across five different map configurations (30×30 to 120×50):
Metric Improvement Over Traditional A*Search nodes76.4% average reduction Total turning angle71.7% average reduction Hazardous paths Completely eliminated Path length1–5% longer but significantly safer Algorithm time Consistently faster than A*, Dijkstra, and RRT
These are not marginal improvements — they represent a fundamental enhancement of how A* operates in complex environments.
Paper B — Broader Context: Where Adaptive Search Leads
Connecting A* Intelligence to Agentic AI
Paper B reveals that the principles behind Paper A's improvements — adaptive planning, contextual reasoning, memory-guided decisions, and self-improvement — are exactly the principles powering Agentic AI. Banda et al. identify six core components in every agentic system, each mapping to Paper A's innovations:
Agentic Component Paper A's A* Equivalent Perception Grid map obstacle detection Planning & Reasoning A* with adaptive weighting Memory Open/closed lists + reward field Execution Path generation and expansion Reflection Two-pass path optimization Or chestration RBF weight controlling search behavior
This mapping reveals something remarkable: Paper A's improved A* already exhibits proto-agentic behavior. It adapts its strategy based on context (RBF weighting), uses pre-computed environmental knowledge to guide decisions (reward field), and refines its output through iterative self-evaluation (secondary optimization). These are the defining characteristics of agentic intelligence.
Course Connection: These components extend the classic sensors → decision-making → actuators agent architecture with persistent memory, reflection, and multi-agent orchestration. The five architectural patterns identified — Re Act loops (like A*'s evaluate-expand cycle), Supervisor/Hierarchical (like goal decomposition), Hybrid Reactive-Deliberative (like our textbook's hybrid agents), BDI (beliefs/desires/intentions), and Layered Neuro-Symbolic — map directly to agent types we studied. Frameworks like Auto GPT (goal-based agent), Auto Gen (multi-agent coordination), and Meta GPT (agent societies) implement these patterns at scale across healthcare, finance, transportation, and scientific research — proving course concepts are deployed industry-wide.
Frameworks and Real-World Scale
The paper evaluates nine LLM-based frameworks (Lang Chain, Auto GPT, Baby AGI, Auto Gen, Meta GPT, CAMEL, Super AGI, and others) and catalogs applications across 13+ domains including healthcare, finance, transportation, manufacturing, education, and scientific research. Auto GPT, for instance, takes high-level goals, decomposes them into sub-tasks, executes them using external tools, and refines its approach based on outcomes — essentially performing the same plan-execute-evaluate-refine cycle that Paper A's improved A* performs, but at the scale of entire business workflows rather than individual path segments.
Critical Insight: What Deep Analysis Revealed
On Paper A: Surface reading suggests "faster A*." Deep analysis reveals something profound: the RBF adaptive weight dissolves the boundary between algorithms we studied as separate entities. The algorithm continuously morphs between greedy and optimal search based on spatial context — it is not a fixed procedure but an adaptive reasoning strategy. The reward field similarly transforms A* from topologically blind to topologically aware. Together, these innovations give A* proto-agentic behavior: it adapts strategy to context, escapes traps using environmental knowledge, and refines outputs through self-evaluation.
On Paper B: Initially appearing as "just a survey," careful analysis reveals its contribution is taxonomic — creating structured vocabulary for a field that lacked one. The Venn diagram showing agentic AI as a convergence of LLMs, reinforcement learning, and multi-agent systems crystallized that agentic AI is not a single technology but an intersection of every AI subfield our course covered.
The Bridge: Both papers share identical limitations — Paper A cannot handle dynamic obstacles or 3D environments; Paper B acknowledges agentic AI struggles with reliability, coordination failures, and accountability. From different scales, both point toward the same frontier: building AI that is reliably intelligent in unpredictable conditions. The progression is clear: the adaptive intelligence we build into individual algorithms scales into the autonomous behavior defining agentic systems.
What This Means for Our Course
The intellectual progression these papers reveal — from static algorithms to adaptive search to autonomous agents — mirrors the arc of our entire course. We started with uninformed search (BFS, DFS), added heuristic guidance (A*), introduced agent architectures (reflex, goal-based, utility-based, learning), and discussed multi-agent coordination. Paper A sits at the transition point between "smarter algorithms" and "intelligent behavior." Paper B shows where that transition leads. Together, they demonstrate that the algorithms we learn are not just exam material — they are the building blocks of autonomous systems reshaping every industry.
Conclusion
Wang et al.'s improved A* demonstrates that classical algorithms become dramatically more powerful through adaptive weighting, reward-based guidance, and iterative self-improvement — achieving 76.4% fewer nodes and eliminating hazardous paths entirely. Bandi et al.'s survey shows these same principles — planning, memory, reflection, and goal pursuit — scaling into fully autonomous systems across every major industry. For AI students, the message is unmistakable: the algorithms we learn are the building blocks of autonomous systems reshaping our world. Understanding how A* works gives us the foundation; understanding why agentic AI matters gives us the direction. Intelligence is not just about optimal search — it is about adaptive decision-making, contextual reasoning, and systems that learn to think for themselves.
> m-zaid-3-5.hashnode.dev
Top comments (0)