Long-term symbolic learning

What are the characteristics of long-term learning? We investigated the characteristics of long-term, symbolic learning using the Soar and ACT-R cognitive architectures running cognitive models of two simple tasks. Long sequences of problems were run collecting data to answer fundamental questions about long-term, symbolic learning. We examined whether symbolic learning continues indefinitely, how the learned knowledge is used, and whether computational performance degrades over the long term. We report three findings. First, in both systems, symbolic learning eventually stopped. Second, learned knowledge was used differently in different stages but the resulting production knowledge was used uniformly. Finally, both Soar and ACT-R do eventually suffer from degraded computational performance with long-term continuous learning. We also discuss ACT-R implementation and theoretic causes of ACT-R’s computational performance problems and settings that appear to avoid the performance problems in ACT-R.