Aristotle's Topics gave philosophy one picture of reasoning: dialectical and staged, where a conclusion can be drawn, contested, and revised as anargument proceeds. Much of the logic that followed pictured consequence differently, fixed and all-at-once: adding premises only adds conclusions, nevertaking one back. Yet real reasoning is not like that: learning more can force an agent to give up what she had concluded. Non-monotonic logic capturesexactly this. ADAPT starts from a simple question: what may an agent rationally conclude when her time, memory, computational or cognitive power islimited, and suspension of judgement is not possible? The question is old in philosophy and logic, and found to be urgent again in bounded rationality and inthe study of artificial reasoners that trade depth for speed; yet logic still lacks an account of it.
Non-monotonic logic indeed has so far grown along a single axis: how an agent's conclusions change as her information changes. ADAPT adds a secondaxis. With information held fixed, conclusions can also change as the agent reasons further; what she accepts after little reasoning, she may give up aftermore. A conclusion then depends on two things, what the agent knows and how far she reasons from it. The first kind of variation we call informational non-monotonicity, the familiar one; the second, bounded non-monotonicity. The project's object is a family of non-monotonic consequence relations indexed bydepth. ADAPT's originality is foundational, not incremental: the question takes a line to state, yet is neither simple to answer nor shallow in consequence.
ADAPT closes a loop between philosophy and formalism. Philosophical analysis fixes what inferential depth is. Formal work then states rationality postulatesfor how commitments appear, are defeated, return, and finally settle as depth grows, and tests them in six paradigms, namely: answer-set programming,conditional and typicality logics, argumentation, weighted and algebraic structures, ontologies, and dynamic-epistemic logic. The method is what the loopdoes with the result: a postulate that holds across most paradigms is a candidate for a common core, while one that fails sends a precise question back tophilosophy: the paradigm needs a different notion of depth, or the postulate was too strong.
The central claim, that the two axes are genuinely independent, is no conjecture: it was already established in a few paradigms by members of theconsortium; ADAPT extends, generalise and systematises it. The result is a foundational account of agents that are rational not only by what they know, but by how far they can reason from it.
Non-monotonic logic indeed has so far grown along a single axis: how an agent's conclusions change as her information changes. ADAPT adds a secondaxis. With information held fixed, conclusions can also change as the agent reasons further; what she accepts after little reasoning, she may give up aftermore. A conclusion then depends on two things, what the agent knows and how far she reasons from it. The first kind of variation we call informational non-monotonicity, the familiar one; the second, bounded non-monotonicity. The project's object is a family of non-monotonic consequence relations indexed bydepth. ADAPT's originality is foundational, not incremental: the question takes a line to state, yet is neither simple to answer nor shallow in consequence.
ADAPT closes a loop between philosophy and formalism. Philosophical analysis fixes what inferential depth is. Formal work then states rationality postulatesfor how commitments appear, are defeated, return, and finally settle as depth grows, and tests them in six paradigms, namely: answer-set programming,conditional and typicality logics, argumentation, weighted and algebraic structures, ontologies, and dynamic-epistemic logic. The method is what the loopdoes with the result: a postulate that holds across most paradigms is a candidate for a common core, while one that fails sends a precise question back tophilosophy: the paradigm needs a different notion of depth, or the postulate was too strong.
The central claim, that the two axes are genuinely independent, is no conjecture: it was already established in a few paradigms by members of theconsortium; ADAPT extends, generalise and systematises it. The result is a foundational account of agents that are rational not only by what they know, but by how far they can reason from it.