Tamal Biswas

Psychometric modeling of decision making via game play

We build a model for the kind of decision making involved in games of strategy such as chess, making it abstract enough to remove essentially all game-specific contingency, and compare it to known psychometric models of test taking, item response, and performance assessment. Decisions are modeled in terms of fallible agents Z faced with possible actions ai whose utilities Ui=U (ai) are not fully apparent. The three main goals of the model are prediction, meaning to infer probabilities Pi for Z to choose ai; intrinsic rating, meaning to assess the skill of a persons actual choices ait over various test items t; and simulation of the distribution of choices by an agent with a specified skill set. We describe and train the model on large data from chess tournament games of different ranks of players, and exemplify its accuracy by applying it to give intrinsic ratings for world championship matches.

Measuring Level-K Reasoning, Satisficing, and Human Error in Game-Play Data

Inferences about structured patterns in human decision making have been drawn from medium-scale simulated competitions with human subjects. The concepts analyzed in these studies include level-k thinking, satisficing, and other human error tendencies. These concepts can be mapped via a natural depth of search metric into the domain of chess, where copious data is available from hundreds of thousands of games by players of a wide range of precisely known skill levels in real competitions. The games are analyzed by strong chess programs to produce authoritative utility values for move decision options by progressive deepening of search. Our experiments show a significant relationship between the formulations of level-k thinking and the skill level of players. Notably, the players are distinguished solely on moves where they erred -- according to the average depth level at which their errors are exposed by the authoritative analysis. Our results also indicate that the decisions are often independent of tail assumptions on higher-order beliefs. Further, we observe changes in this relationship in different contexts, such as minimal versus acute time pressure. We try to relate satisficing to insufficient level of reasoning and answer numerically the question, why do humans blunder?

A Comparative Review of Skill Assessment: Performance, Prediction and Profiling

The assessment of chess players is both an increasingly attractive opportunity and an unfortunate necessity. The chess community needs to limit potential reputational damage by inhibiting cheating and unjustified accusations of cheating: there has been a recent rise in both. A number of counter-intuitive discoveries have been made by benchmarking the intrinsic merit of players’ moves: these call for further investigation. Is Capablanca actually, objectively the most accurate World Champion? Has ELO rating inflation not taken place? Stimulated by FIDE/ACP, we revisit the fundamentals of the subject to advance a framework suitable for improved standards of computational experiment and more precise results. Other games and domains look to chess as demonstrator of good practice, including the rating of professionals making high-value decisions under pressure, personnel evaluation by Multichoice Assessment and the organization of crowd-sourcing in citizen science projects. The ‘3P’ themes of performance, prediction and profiling pervade all these domains.

Building Long Term Trust in Vehicular Networks

In vehicular networks (VN), the response time is critical, whereas, an autonomous and efficient way of preventing hazardous situations on roads plays an important role in the successful deployment of the system. Trust or reputation models often need to be integrated with inter-vehicle communication protocols in use to avoid selfish or malicious behavior by the vehicles exploiting the system. Existing trust and reputation models for VNs lack the capability of fast and accurate trust management suitable for the ephemeral association of vehicles. In this paper, we design an augmented trust model for VNs by assigning each vehicle a long term trust value. Our model eliminates the overhead of repeated bootstrapping and ensures accountability of the vehicles for incident reporting and other critical actions. The two main features of our framework, viz. a three- party authentication and privacy protocol, and a trust propagation model, both of which are crucial for a successful deployment of VNs, can also be used in any generic security application.

Approximation of function evaluation over sequence arguments via specialized data structures

Abstract This paper proposes strategies for maintaining a database of computational results of functions f on sequence arguments x → , where x → is sorted in non-decreasing order and f ( x → ) has greatest dependence on the first few terms of x → . This scenario applies also to symmetric functions f, where the partial derivatives approach zero as the corresponding component value increases. The goal is to pre-compute exact values f ( u → ) on a tight enough net of sequence arguments, so that given any other sequence x → , a neighboring sequence u → in the net giving a close approximation can be efficiently found. Our scheme avoids pre-computing the more-numerous partial-derivative values. It employs a new data structure that combines ideas of a trie and an array implementation of a heap, representing grid values compactly in the array, yet still allowing access by a single index lookup rather than pointer jumping. We demonstrate good size/approximation performance in a natural application.

Efficient Memoization for Approximate Function Evaluation over Sequence Arguments

This paper proposes strategies for maintaining a database of computational results of functions f on sequence arguments x, where x is sorted in non-decreasing order and f(x) has greatest dependence on the first few terms of x. This scenario applies also to symmetric functions f, where the partial derivatives approach zero as the corresponding component value increases. The goal is to pre-compute exact values f(u) on a tight enough net of sequence arguments, so that given any other sequence x, a neighboring sequence u in the net giving a close approximation can be efficiently found. Our scheme avoids pre-computing the more-numerous partial-derivative values. It employs a new data structure that combines ideas of a trie and an array implementation of a heap, representing grid values compactly in the array, yet still allowing access by a single index lookup rather than pointer jumping. We demonstrate good size/approximation performance in a natural application.