Jason Induction of Logical Decision Trees: A Learning Library and Its Application to Commitment

Abstract

National audienceThis paper presents JILDT (Jason Induction of Logical Decision Trees), a library that defines two learning agent classes for Jason, the well known java-based implementation of AgentSpeak(L). Agents defined as instances of JILDT can learn about their reasons to adopt intentions performing first-order induction of decision trees. A set of plans and actions are defined in the library for collecting training examples of executed intentions, labeling them as succeeded or failed executions, computing the target language for the induction, and using the induced trees to modify accordingly the plans of the learning agents. The library is tested studying commitment: A simple problem in a world of blocks is used to compare the behavior of a default Jason agent that does not reconsider his intentions, unless they fail; a learning agent that reconsiders when to adopt intentions by experience; and a single-minded agent that also drops intentions when this is rational. Results are very promissory for both, justifying a formal theory of single-mind commitment based on learning, as well as enhancing the adopted inductive process

Agents that Argue and Explain Classifications of Retinal Conditions

“…Hernandez et al have used Jason Induction of Logical Decision Trees (JILDT) [12] to create logical decision trees using AgentSpeak plans. The focus is on developing agents that learn while they are executing plans.…”

Decision as choice of potential intentions

“…One benefit of this research, which attempts to build a formal connection between agent mental attitudes and the meta-cognition concepts as used in RPD, is that it allows some well-founded BDI social learning models [28] to be embedded into the RPD framework. One example is the inductive decision learning approach [15], which allows agents to learn over plans (i.e., recipes) within the BDI framework. A virtue of inductive decision learning is that it allows agents to de-rive the best explanations of why some of their decisions succeeded or failed.…”

“…In so doing, agent A i actually takes a bias: the recipe that has been applied most successfully is dequeued (line 3) and considered first. The potential intentions in Υ P i are then evaluated one by one (lines [5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21]. This process repeats until a workable solution is obtained (lines 12-14), or Υ P i becomes empty, which indicates that no solution can be constructed and the situation ought to be further diagnosed (refer to Section 6).…”

“…In such a case, the evaluation process is done, and all the assigned agents should proceed to implement α. -If A i judges that α is not workable but adjustable, it will invoke CollectInsights() to collect insights from the other peers (lines [15][16]. The insights in ω will be used to reify ρ in the next round of COA construction.…”

“…the completion of a decision task). There are three stages for each round of monitoring: multi-agent situation awareness (lines 3-5), anomaly detection (lines [6][7][8][9][10][11][12][13][14][15][16], and corrective actions (lines [17][18][19][20][21][22].…”