MGPO discovers significantly more resource-rational planning strategies than baseline methods in partially observable environments¶

Claim · All claims
Evidence · 1 study · causal r3 · q1
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q1 i? In simulation benchmarks across four partially observable environments, MGPO achieved higher resource-rationality scores than the meta-greedy policy and all PO-UCT configurations. → Heindrich 2025

Evidence¶

Heindrich 2025¶

Heindrich, L., Consul, S., & Lieder, F. (2025). An Intelligent Tutor for Planning in Large Partially Observable Environments. International Journal of Artificial Intelligence in Education, 35, 3095–3127. https://doi.org/10.1007/s40593-025-00493-7

q1 · i? · causal · r3

Simulation benchmark comparing MGPO against meta-greedy and PO-UCT baselines across four environments of 37–91 nodes, evaluated on 5000 random instances per condition. A three-way ANOVA (F(5, 239952) = 32682.5, p < .001) confirmed algorithm differences, with "MGPO performs significantly better" on post-hoc tests.

"A Tukey-HSD post-hoc comparison showed that MGPO performs significantly better than the meta-greedy policy ( p = .001), the PO-UCT algorithm with 5000 steps ( p = .008), and all PO-UCT algorithm configurations with less than 5000 steps (all p <. 001)."

Discussion¶

Related Claims¶

• MGPO selects planning computations in approximately 0.03 seconds, over 100 times faster than baseline methods — related
• A preregistered experiment with 330 participants demonstrated the intelligent tutor is highly effective at improving planning in partially observable environments — related