Adaptive Meta-Reasoning for Algorithm Selection in Cognitive Agent Architectures

Autonomous agents operating in dynamic environments face a fundamental challenge: no single algorithm consistently performs best across all conditions. Whether navigating physical spaces, allocating computational resources, or selecting reasoning strategies, agents that commit to a fixed algorithmic approach at design time may risk underperforming when runtime conditions deviate from initial assumptions. This thesis addresses the problem of enabling autonomous agents to adaptively select the best-performing algorithm at runtime while maintaining the transparency of their deliberative reasoning, using grid-based pathfinding as the experimental domain in which the proposed approach is instantiated and evaluated. To address this challenge, first, we conduct a comprehensive empirical analysis of six heuristic search algorithms across systematically varied grid-based pathfinding environments, demonstrating that algorithm performance is highly sensitive to environmental characteristics such as grid size, start-goal distance, and obstacle density. Second, we investigate whether this performance variability is learnable, training Random Forest classifiers on environmental features to predict optimal algorithm choices for static problem instances. Models achieved prediction accuracies of up to 84\% for memory consumption minimization, confirming that algorithm selection is feasible as a supervised learning problem, while also exposing a fundamental limitation: static classifiers cannot revise their choices once execution begins and environmental conditions evolve. To overcome this limitation, we propose a novel agent architecture that embeds adaptive algorithm selection within the agent's deliberation cycle rather than treating it as an external preprocessing step. The architecture extends the Belief-Desire-Intention (BDI) framework with a reinforcement learning meta-reasoner, implemented as a Deep Q-Network, that continuously monitors the agent's belief--intention state and dynamically selects among planning algorithms at runtime. Algorithm selection is formalized as a sequential decision-making problem extending Rice's classical framework to dynamic settings, where states encode environmental observations and deliberative context, actions correspond to algorithm choices, and rewards measure the quality and efficiency of planning outcomes. Experimental evaluation in dynamic pathfinding environments with moving obstacles demonstrates that the trained meta-reasoner achieves a success rate of 99.43\% and substantially lower penalties compared to all fixed-algorithm baselines, with the weakest baseline reaching only 88.83\% success at over ten times the penalty cost. The agent learns a behaviourally interpretable and context-sensitive strategy, favouring deeper lookahead in sparse environments and switching to faster replanning as obstacle density increases. These results confirm that integrating learned meta-level reasoning into a cognitive architecture yields tangible performance advantages over static approaches while preserving the transparency and goal-directed structure of the BDI deliberation layer.