|
| Relational Responding and Artificial Agents |
| Saturday, May 25, 2024 |
| 12:00 PM–12:50 PM |
| Convention Center, 200 Level, 204 C |
| Area: EAB/PCH; Domain: Basic Research |
| Chair: Martin Finn (Ghent University) |
| Abstract: There is increasing recognition that the ability to respond arbitrarily according to various kinds of stimulus relations (Penn et al., 2008) and to derive new stimulus relations based on previously learned stimulus relations is central to human cognitive abilities (McLoughlin et al., 2020). Simultaneously, there is extraordinary growth in research on artificial intelligence and producing machines that learn and think like humans (Lake et al., 2017). However, little work has brought these ideas and aims together. This symposium presents efforts at producing relational responding with artificial agents that draw heavily on behavior-analytic work on equivalence relations (Sidman, 1994) and arbitrarily applicable relational responding (Hayes et al., 2001). Paper 1 describes a reinforcement learning agent capable of completing the procedures described in Steele & Hayes (1991) in a manner highly comparable to humans exposed to the same procedure. Paper 2 introduces the Non-axiomatic reasoning system (NARS; Wang, 2013) and demonstrates its capacity for learning relations in a variety of environments and compares NARS and reinforcement learning and performance of agents based on these systems. Paper 3 demonstrates that when given training of the kind suggested by relational frame theory the NARS demonstrates relational responding in accordance with symmetry. |
| Instruction Level: Intermediate |
| Keyword(s): artificial intelligence, relational responding |
| |
A Reinforcement Learning Model of Arbitrarily Applicable Relational Responding in the Seminal Steele & Hayes (1991) Procedure |
| MATTHIAS RAEMAEKERS (Ghent University), Martin Finn (Ghent University), Jan De Houwer (Ghent University) |
| Abstract: Relational Frame Theory is but one of many theoretical perspectives that consider relational reasoning a central aspect of human higher cognition. Despite spectacular advances in recent years, state-of-the-art artificial intelligence is still challenged by the uniquely human capacity for arbitrarily applicable relational responding (AARR), responding to relations between event not defined by their physical properties, but by the context. Furthermore, the mainstream reliance on a (big) data-driven approach to train such models highlights differences with the speed and flexibility by which humans develop this ability. In a seminal publication, Steele & Hayes (1991) demonstrated the arbitrary, flexible and generative nature of AARR. We present simulations of an artificial agent exposed to this procedure, and show that its hardcoded relational learning mechanisms are sufficient to reliably solve derived relational responding test problems. Furthermore, we fit the model to behavioral data from human participants (N = 175), showing the model’s capacity to predict participants responses. This represents the first step towards the development of a computational model of AARR. Such a model would allow us to simulate study the development of AARR with less practical limitations, and could be used to test the effects of relational training interventions. |
| |
Reasoning-Learning Systems for Adaptive Autonomous Agents Based on Non-Axiomatic Reasoning System Theory |
| PATRICK HAMMER (Stockholm University), Tony Lofthouse (Stockholm University), Robert Johansson (Stockholm University) |
| Abstract: We will present Non-Axiomatic Reasoning System (NARS) for advanced reasoning and decision making in adaptive intelligent agents. We present this reasoning system particularly as an alternative formal model to Reinforcement Learning. The system works under the Assumption of Insufficient Knowledge and Resources which demands both open-ended adaptation and real-time operating capabilities under strict computational resource constraints, which imposes strong constraints on both its memory and inference control mechanism which we will briefly describe. Then, we show data-efficient adaptation of a NARS-controlled agent in a Non-Markovian domain, where we demonstrate inherent advantages over a Q-Learner, in addition to being significantly less dependent on hyperparameter tuning. Furthermore, we present the system’s ability to plan ahead sequences of actions which have previously not been experienced in that particular order, which is particularly important when novel outcomes should be accomplished by the agent. This capability is particularly useful in combination with learning cause-effect and other relations directly from streams of input events, as it can make agents more autonomous and less dependent on human-provided knowledge, while at the same time being able to utilize it effectively when available. |
| |
Symmetry With the Non-Axiomatic Reasoning System |
| ROBERT JOHANSSON (Stockholm University), Tony Lofthouse (Stockholm University), Patrick Hammer (Stockholm University) |
| Abstract: In this talk, we will demonstrate how the Non-Axiomatic Reasoning System (NARS; Wang, 2013) can learn symmetry within the context of matching-to-sample via multiple-exemplar training. NARS is a adaptive reasoning system that can operate with insufficient knowledge and resources. This exemplifies our research approach where we focus on evaluating psychological abilities of NARS, with increasing complexity, using methods from behavioral psychology. Our previous research has involved demonstrating that NARS can learn through classical and operant conditioning, and also that it can learn generalized identity matching. This particular research aims to demonstrate how NARS could learnin symmetry. First, we show how the matching-to-sample is encoded with NARS and how NARS is able to learn from repeated interactions with the procedure. Then, we will explain how the different mechanisms of NARS combine to enable the system to learn symmetry through multiple-exemplar training. Future directions of this work, and the implications of a system like NARS for behavioral psychology research will also be discussed. |
|
| |