A METHOD AND SYSTEM FOR PERFORMING IMAGE BASED ANALYSIS OF STATE INFORMATION OF GRAPHICAL ELEMENTS OF A GRAPHICAL USER INTERFACE

FIELD OF THE INVENTION

The present invention relates generally to the field of analysing state information of graphical user elements of a graphical user interface (GUI), and more specifically, to image based analysis of state information of graphical tabs of such graphical user interface (GUI).

BACKGROUND OF THE INVENTION

Artificial Intelligence (Al) and machine learning have been advancing the gaming domain over the years, primarily with the objective of developing automated agents or bots to compete with humans.

With each milestone advancement in Al technologies, computers have exceeded human experts in several games such as backgammon, checkers, chess, Jeopardy!, Atari video games, and SO ON.

This presents a technical problem as it is critical to ensure that there is no interference in the game play via any kind of automated decision making (beyond e.g. from the random dealing of cards), to keep it fair and real, which is the main value proposition of such games.

To achieve said objective it is required to understand, benchmark, and profile the individual players with respect to their playing behavior, strategies, playing patterns and the longitudinal evolution of game play, (i.e., changes over time).

Multi-player skill-based card games contain a large amount of game state information which, when mined, yield valuable information about the players and the game. Such analyses are used for understanding the behavior of players especially from the perspective of how the player reacts to a game state. Examples of such behaviors include but not limited to determining if the player is conservative or a risk-taker or skilled, etc. Benchmarking and profiling players' actions and behaviours in different game state improves understanding of player motives which eventually helps in identifying automated agents or bots. For example, one can possibly identify if a player has become more aggressive or disengaged after getting bad cards for consecutive games.

In the distributed network environment the various graphical user interface (GUI) elements contain a large amount of state information which, when extracted, yields valuable information about the behaviour of the users interacting with such graphical user interface (GUI) elements.

For Example, the behaviours may include determining if the user is a skilled user or not.

Such a determination may be based on from the perspective of how the user reacts to various states of a Graphical User Interface (Element).

However, understanding such behaviour of users from the GUI interaction patterns requires complex analysis of data from different granularities and viewpoints. In known and traditional machine learning systems, such information is derived in the form of features for the statistical models to learn and predict correctly. Identification of these features often require extensive domain understanding and cumbersome manual interpretation.

However, the existing systems face the daunting challenge of representing raw GUI elements state information in a way that can be effectively interpreted by humans and machines.

Further, understanding behaviors of players from card game play patterns and benchmarking/profile players involves technical challenges including, e.g.

(1)

It typically req uires complex analysis of data at different granularities and time scales. The state space for such a domain, where each player is dealt several cards, and often, more than one deck of cards is used, and each game takes many rounds to complete, becomes enormous, making traditional machine learning systems unwieldy.

(2)

The features for classical machine learning models are game-specific and have to be derived based on intricate knowledge ofthe game rules and nuances.

(3)

A typical benchmarkingframework needs to capture the state from the players perspective relative to one another in the game. For example, the quality of a set of cards dealt to a player needs to be qua ntified in a continuous domain instead of simply classifyin them to be good or bad. This is because a player with good cards can actually be playing with another player with even better cards.

Accordingly, although a state information is imperfect from a players' perspective, a benchmarking framework needs to allow flexibility to capture such relativity.

(4)

Benchmarking of playing patterns is fundamentally different from various player rating systems in different games (e.g., ELO rating in chess [14]).

This is because reaction to game states in multiplayer card games is often personal. For example, in the typical game of cards one important decision for a player is whether to play or drop based on the cards he was dealt. For the corner cases, i.e., clearly very good or bad set of cards the decision is often straightforward to play or drop, respectively.

However, most of the sets typically fall in intermediate category. While a conservative player may drop such hands, a more aggressive player may play with those and may possibly end up winning depending on the cards available to other players, and also how thegame evolves. Both kind of players may end up with very similar ratings, based on any established rating schemes which look at their overall standing across all the players and all the games in the system.

Hence, we strive for amore fine-grained observation of player behaviors.

Therefore there is a long felt need to develop a novel structure of representing the GUI elements state information.

The present invention proposes a technical solution for representing the GUI elements state information in the form of a rich multi-dimensional image. The structural representation proposed by the present invention encapsulates the relationship among GUI elements such as tabs and also the meta data relevant to such GUI elements.

DESCRIPTION OF THE INVENTION

The present invention is directed to analysing and representing the state information of GUI elements in a distributed network environment. The invention is a combination of hardware and software elements that operates in a client server based network environment.

In an embodiment of the invention a fine-grained analysis of game play and action data is performed to gain automated insights on playing behaviour e.g. So that targeted services and assistance can be provided for improved and personalized experience in card gaming platforms.

In an embodiment the present invention usesa unique and extensibleformatto representthe stateofanycard gameas an image and show that Convolutional Neural Network (CNN) models can learn these images efficiently.

In an embodiment CN Ns are trained to gather deep insights of the game, at different scales (player-level, game-level and even segment-level), and also at different time granularities (next move, game end, and even over days) and models trained on such representation for millions of game records, are used for learning critical game decisions, benchmarking playing patterns, characterizing player personas.

In an embodiment of the invention the images are fed to deep image learning models, such as Convolutional Neural Networks (CNN) for learning user interaction patterns.

To overcome the disadvantages of existing methods and systems, the present invention proposes a Game Action Information Miner (GAIM), a Convolutional Neural Network based framework for card game analytics.

They provide a novel representation of the card game state as an image (three dimensional array), which encapsulates all pertinent information in a concise yet extensible and generic manner.

The Convolutional Neural Network (CNN) model in line with present invention is trained to estimate (and subsequently predict) different game checkpoints of interest, directly from the raw state representation, thereby eliminating the need to derive game-specific and goal-specific features. This advantageously helps in benchmarking players, with respect to true state behavior. This in turn is quite valuable for player and segment profiling, and identifying the behavioural characteristics of players (for example, aggressive vs. conservative or new user who needs assistance, etc.).

The present invention proposes an information mining framework, GAIM, which analyses the game action data by: (i) transforming the states into a unique, generic, and extensible image representation; and subsequently, (ii) gathering deep insights of the game, at different scales (player-level, game-level and even segment-level), and also at different time granularities (next move, game end, and even over days) in skill-based card games.

The purpose of the GAIM framework is to enable deep analysis of the actions taken by the players during the course of a game, with respect to how they react to different situations and state evolutions. The results from such analysis help in deriving insights about players which can be used to enrich the analytical models

This is illustrated in Figure 2 which depicts that the GAIM layer sits in between the raw data bases and the analytical models that consume this data, augmenting the raw data with additional derived information.

As discussed before, the main information in a game at any given time, includes (i) the players' hands, (ii) wild card joker, (iii) the closed stack cards, and the (iv) discard stack cards. In GAIM, we represent a player's hand as a 4x14x3 array, where,

- The first dimension is for the 4 suits (in order of F, F, ¨, *)

- The second dimension represents the face values of the cards (in the order A, 2, 3, . . . , 10, J, Q, K, A). It may be noted that A is repeated in the first and fourteenth columns in order to indicate that it can take either place in the sequential ordering of cards, i.e., it allows both {A, 2, 3} and {Q, K, A} to be recognized as sequences.

- The third dimension consists of a 3-tuple where the first component is the count of the card corresponding to the grid cell. For example, if the player has one 4F, then the value of the first component in cell 1, 4 will be 1. Since two decks are used, the possible values are 0, 1 or 2.

The second component is used to indicate special properties specific to the corresponding card. For instance, we use it to represent if the card is a wild card joker. The third is used to represent properties that are common to the entire hand. For instance, we use it to denote the total number of jokers in the hand.

This also allows for the printed joker card (which does not have a place in the 4x14 array) to be counted.

Furthermore, it is structurally better to position the joker (and other information that are common across all cells) as a separate plane rather than as an additional cell because, semantically this information is uniform across all cells (and as we see in Section 5, it helps the convolution window learn the role of the joker in the game. For example, suppose a hand consists of the following cards: (9*, 2 , 8 , 5^, 7+, 6F, 6+, 2F, Q 4 , Joker

, 94 }, and the joker card is 9F (which means that 9 of any suit can be used as wild card joker). The array representations in each of the three planes for this hand is shown in Figure 3.

The 3-tuple can be taken as the RGB components, resulting in an image representation as shown in Figure 4WVe can now visualize the ha nds as an image, wherein pure sequences a re seen as horizontal consecutive blocks and sets are seen as vertical blocks; multiple cards of the same rank and suit result in a brighter red block; jokers emerge as a greenish block because it corresponds to the green- component; more number of jokers result in a blue-tinge in the background since the blue-component is uniform across the entire image.

Although this representation is sufficient for ca pturing all the information we need for the analyses discussed within the scope of this work, it may be noted that each of the components in the 3-tuple are capable of more information, sincethey each have 8-bit representations. For example, we can use the second component to indicate whether a cell is a joker card and/or the top card in the discard stack that a player sees in his turn, then the second component can be used, with bits 00, 01, lOor 11. It may also be noted that the same representation can be used to represent the closed stack and discard states. Also, this representation can be applied to other games as well.

Further, FIGURE 4 depicts a convolution neural network architecture that is capable of learning the images.

One of the objective is to learn horizontal sequences and vertical sets of size 3 or 4 cards (cells). Furthermore, the CNN must learn that the jokers play a role across the entire hand. Accordingly, the architecture of the CNN is as shown in Figure 5, consisting of three 2D convolution layers of size 4x4 and depth 64, 128 and 256, respectively, followed by an Average Pooling layer of size 2x2 and a Dropout layer, the output of which is flattened and input to a dense network of 512 nodes.

The activation functions for the convolution and dense layers are RELU and tanh, respectively. We refer to this architecture as the Fland Image Convolution Unit (HICU), and consider this as the basic building block using which models with specific objectives can be built.

As the first and the main use case for the GAIM framework, we build a model for hand quality estimation. When the cards are dealt to a player, he needs to respond to the hand depending on whether it is a good, bad or mediocre hand. As mentioned in Section 3, the decision to drop or play in the first round is one of the most important decisions a player makes. In case of a clearly strong hand (for example, one that has a pure sequence already) or a clearly bad hand (for example, one with no jokers, and cards that are far apart), it is an easy decision to play or drop, respectively, but with mediocre hands, the decision is much more difficult and would depend on the skill level as well as behavioral tendencies of a player. We model the first drop decision of skilled players and the probability of drop for a given hand which is output by the model provides an effective, continuous-ranged proxy for hand quality.

This model of present invention (i) learns the nuances of the game accurately even though there is no game-logic related feature input to the model, (ii) is accurate, and most mispredictions can be attributed to either player mistakes or intermediate (hard to decide), (iii) can be used as a benchmark model to gather valuable behavioral insights about players by analyzing player response to hand quality.

For building this model, present invention selects a set of skilled players from a platform. A skilled player may be defined as one who has played at least 500 games and has an average differential score of < -5. The average differential score of a player is defined as

A higher value of negative score implies a more skilled player. The lower bound on the number of games played ensures that there is no luck involved in the computation of the skill score.

These filters resulted in a skilled-player base of 2394 users out of the total player base of 282,102 users over a period of 2 months. The initial hands of all the games played by these users is taken as input, and the binary decision of drop (I) or play (0) istaken asthe ground truth label. The dataset contained 2,944,242 records.

The architecture of the model is shown in Figure 6 where the Hand Image Convolution Unit of Figure 5js then input to a drop out layer (20%), and then a softmax layer is used for binary classification, with categorical cross entropy as loss function.

The model of the present invention is trained for 100 epochs, with a stochasticgradient descent optimizer and loss function of categorical cross entropy. We refer to this CNN (HICU) based hand quality estimation model as HQE in the rest of this paper.

The first drop decision of a player depends significantly on the number of players in a game. The tendency to drop is higher in a 6-player game than a 2-player game, since there is higher chance of not getting the cards needed to win.

Therefore separate models are trained for 2-, 3-, 4-, 5- and 6-player games. In each case, the dataset is split into training, test and validation sets in the proportion of 90%, 5%and 5%, respectively by way of example.

Also, the present invention evaluates hand quality estimation model with respect to not only how well the model learns the ground truth that it was trained on, but also in terms of how well the game logic and nuances are learned (including our primary objective of estimating hand quality).

The performance metrics of the 2-player model on the validation set are shown in TABLE 1, along with the ROC curve, Precision- Recall curve (PRC) in FIGURE 7.

It may be noted that in the rest of the evaluation, we optimize the classification threshold for the Break Even Point, i.e., the threshold at which both precision and recall are highest.

This approach is preferred (to other metrics such as F - 1 score because), as one of the objective of present invention is to effectively identify the good, bad and intermediate hands, rather than focusing on correctly predicting the drop class alone.

The HQE model is compared with (i) a Feed Forward neural network, where the array input is flattened and fed to an input layer consisting of 168 nodes, and 3 hidden layers with 1024 nodes each; (ii) Random Forest model and (iii) XGBoost model.

The features that are derived for the XGBoost and RF models are 1 number of jokers, 2 number of pure sequence of length three,

(3) number of pure sequences of length four,

(4) number of pure sequence oflengthtwo (this is also called bits-e.g., 6#, 74), (5) number of sets of length three, (6) number of sets of length four, (7) number of connected cards (e.g., 6 and 8F), (8) minimum distance between two cards of the same suit (i.e., how many cards away from making a pure sequence), (9) hand score, i.e., number of points the hand is worth as described in Section 3 without the cap at 80 points.

Similarly, separate models are trained for 3, 4, 5 and 6 player games. The performance metrics of these models against the validation set are shown inTABLE 2.

Forthe restoftheevaluation 2-player model is used because that is one of the most frequent scenario.

It may be noted here that the conclusions drawn about the two player model can be extended without loss of generalityto > 2-player games as well.

Further, it is evaluated whether the model provides the continuous quality estimation of hands as envisioned, i.e., if the model assigns high probability to poor hands, low probability to good hands and around the threshold value for intermediate hands.

Firstly, hand images are inspected corresponding to the highest and lowest drop probabilities (0.97, 2.76e-07, respectively) in the validation set as shown in FIGURE 8.

A first level of validation is that the probabilities estimate the hand quality, because the least probability hand is already a winning hand; the maximum probability hand has no jokers, and is at least 2 cards away from a pure sequence and hence is clearly a bad hand.

Thereafter, the ending points of the true negatives are compared (i.e., player played when model asked to play) with the false positives (i.e., player played despite the model recommending a drop), to analyze the goodness of the model by retrospectively analyzing how the players fared when they complied with the model versus when they did not.

A one-tailed t-test between the end scores of False Positive and True Negative groups, with the null hypothesis that the mean end scores of FP is not higher than that of TN - there is significant difference in the mean end score of FP games vs. TN games, with a p- value of 3.2e - 05, thereby rejecting the null hypothesis. This means that not following the model's recommendation to drop, yields higher end points (i.e., orse performance) implying they played a bad hand. There is also significant correlation between the drop probability and the corresponding end score (Pearson correlation coefficient of 0.356, p-value of 0.000775 averaged over 20 samples), indicating higher end scores for higher drop probabilities. These evidences point to the fact that the hand quality is being effectively estimated by the drop probabilities.

Still as a further drill-down into the hand quality estimation of model, and to understand why misclassifications occur, a survey of 200 players is utized, where a set of thumb rules were presented and the players were asked if they would categorize the rule as "Play", "Drop", or were "Not Sure".

A summary of thethumb rulesareshown inTABLE 3.

TABLE4 shows the results from our model with respectto these thumb rules.

Thus, most of the misclassified hands were of either Unsure or Drop categories.

Under Play category, 99.1% of misclassified hands are False Negatives, i.e., the player dropped although model recommended play.

This clearly implies user mistakes, possibly due to unfa miliarity of the game.

The average drop probability is similar to the ground truth drop ratio, and is much lesser than the threshold (0.35). Under Drop category, 82.3% of misclassifications are False Positives, this again impliesusermistakes, playingwhen heshould have dropped, possibly due to lackof experience or aggressive playing tendency.

The average drop probability is again similar to the ground truth drop ratio, and higher than the threshold. In the Not Sure category, the predicted and ground truth probabilities are around the threshold, and the FP and FN are similar, indicating that the misclassifications are due to players getting confused with the intermediate hands.

While it may be an impression that the initial hand score or these thumb rules can be used for estimating the hand quality, ratherthan the drop probability.

However, the initial hand score is not a good indicator of the hand quality, because it does not reflect the outcome of the game (in terms of end points, which is essentially the reward/penalty the player gets) regardless of whether the game is dropped or played.

Accordingly in line with present invention a Pearson correlation coefficient between the initial hand scores and the corresponding end scores, and obtained a coefficient value of 0.038 and p-value 0.446, indicating there is no significant correlation between the two metrics).

The thumb rules are not a viable option either because, the survey showed that the players were unsure in majority of the cases (64%), and moreover the survey only provides discrete quality slabs and does not account for number of players in the game. Next, we assess how well the model learns the importance of the joker (both wild card and printed) from our way of hand representation. So in our image, we run the following three experiments:

(i) Reset only the second plane (i.e., the second component of the 3-tuple, referred to as P2) to all zeros and re trained the model

(ii) Reset only the third plane (P3) to all zeros and

(iii) reset both P2 and P3 to zeros

The performance of the models in each of the cases is shown in Table 5.

There is negligible drop in performance when only P2 is removed, but removing P3 results in a 35% drop in precision and recall, removing both P2 and P3 results in 49% drop in precision and recall, substantially validating our structure for hand representation.

Thereafter , as per present invention a randomly selected hand {A+, A , 2*, 4^, 5*, 6 , 6*, 8 , 8*, 8*, 9 , 9F, J F} is taken that does not contain any joker.

Its drop probability is 0.549 (not playable). As a rule of thumb, having a joker improves the quality of hand. So, we changed each of these cards to a joker card (and also introduced a printed JOKER card), and the drop probability was computed in each case (TABLE 6).

We see that the drop probability consistently improves and becomes playable (< 0.35) in all cases.

Since the objective of a typical card game is to minimize the points, a skilled player always tries to discard higher pointcards whenever possible.

Although there is nothing explicitly fed to the model to convey this notion, the model is still of present invention is able to learn this.

For instance, FIGURES 9(A) shows the image of a right-heavy hand (i.e., most of the cards are on the right half of the image, implying many high value cards). The drop probability of this hand is 0.774.

We then flip the hand horizontally, i.e., from left to right, so that the image now becomes left-heavy. The drop probability of the flipped image decreases by 38% to 0.478.

In order to verify this further, the present invention selects 50 right-heavy images with no pure sequences or jokers and computed their drop probabilities from the model, before and after flipping horizontally.

A one-tailed, paired two sample t-test is performed on these two dependent groups, with the null hypothesis that the hand quality does not improve (i.e., drop probability does not reduce) after flip- ping. The pair-wise difference between the samples is shown to be significant with a p - value of 0.00402, rejecting the null hypothesis. Thus, present invention establishes the hand quality estimation model.

Now, present invention describes how to use the hand quality estimation model as a benchmark for player profiling and understanding different player behaviors.

The present invention identifies and evolution in playing patterns of individual users or different player segments, which would help timely interventions to improve retention.

How a player reacts to a hand, given its quality, sheds light on various player characteristics and playing tendencies.

For instance, FIGURE 10 shows the initial hand quality (drop probabilities) of the last few hundred games of two different players.

The green dots are the true negatives, i.e., the player played the hand, and the model also recommended the same; blue dots are the true positives; orange dots are false negatives; pink dots are the false positives.

From this FIGURE 10, we can clearly understand that Player A is a very conservative player, dropping most of the hands he could have played, while Player B is a very aggressive player, playing most of the hands he should have dropped.

To further aid such analyses, present invention further derives relevant metrics (referred to as Drop Metrics), some of which are listed in TABLE 7.

As preliminary validation for these metrics, the present invention computes the Pearson correlation coefficients (i) between the mistake ratio and win ratio (WR) and (ii) between mistake ratio and average end score (ES) for all skilled players in our data set, in order to see if the drop metric has an impact on the game outcome.

There is a significant negative correlation between mistake ratio and win ratio (coefficient = -0.16, p - value = 0.0038), and significant positive correlation between mistake ratio and end score (coefficients.25, p - value = 0.005).

The model did not use any information about the end score or the game result but is trained on only the initial hand and the initial drop labels.

Next, it is submitted that the drop metrics can be used to (i) under- stand the game play characteristics of different player segments, (ii) identify profiles or personas of players based on game play pattern and (iii) improve churn prediction of players.

Using the HQE model as the benchmark, the present invention analyzes that the drop metrics change across different segments of players.

Four different player segments are considered : (1) Platinum Club (PC) players, who are active players who play in the higher bracket of cash games, (2) Poorly skilled (PS) players, who have played at least 300 games and are in the bottom 5% in the differential average score as described above (3) New Users from New Geos (NUNG), who are new players from geographical regions, where a particular card game is not popular and (4) New Users from Old Geos (NUOG), i.e., from regions were a particular card game is well-known and has a large player base.

From each segment, 2000 players are chosen, and the initial hands of their games during a 2 month period were input to the HQE model to obtain the drop probabilities.

TABLE 8 shows the average values of the drop metrics, compared to the skilled set of players (the benchmark values). We observe that the mistake ratios and other metrics for skilled and platinum club players are similar and that of poorly skilled and new users are similar, both of which tally intuitively.

For PC, BDR is the lowest and BPR is the lowest after skilled players, indicating a skilled yet risk-taking tendency. In contrast, for PS, BDR is second lowest and BPR is highest indicating an almost reckless tendency (highest Al). Another observation is that NUOG players tend to be more conservative than NUNG players.

In order to verify that these fine-grained understandings of different segments are statistically significant, we ran a one-way ANOVA on these groups and the p-value was £ 0.0001 for each metric.

Subsequently, a Tukey HSD test [13] is executed for each metric to find out which groups were significantly different (we chose this as a post- hoc test since our objective is to minimize Type I error).

The results of the Tukey HSD test are shown in TABLE 9, where the metrics that are significantly different (with 95% confidence) for each pair of groups are listed in the corresponding cells (the metrics are indexed as in TABLE 7).

It is often useful to profile the players based on game play, i.e., define player personas, in order to provide customized user experiences and player journeys. We show that such personas can be established using the drop metrics.

In order to do that, we randomly selected 2000 active players, generated their drop metrics for the games played over 2 months, and then performed /r-means clustering on the feature set. We set k = 5 based on the knee-point in the scree plot. The cluster centroids and the number of players in each cluster are shown in Table 10. As a retrospectivecrossverification, we also presentsome peripheral, non-game play related information about the players in the clusters (rows 7-11).

Based only on rows 1-6, it can be observed that Cluster 1 consists of players who a re experienced and balanced (lowest end score, not too conservative nor aggressive); Cluster 2 is a set of highly skilled and successful players, who play wisely yet conservatively (low end score, high win ratio, lowest mistake ratio, high bad drop ratio); Cluster 3 can be characterized as reckless, aggressive and unskilled players (lowest DRIH, highest Al, highest ES, lowest WR); Cluster 4 seems to be aggressive but not as reckless and more skilled than as cluster 3 (high WR, low ES, moderate Al); Cluster 0 consists of players who are unskilled, not risk-taking, and seem almost disconnected (which intuitively may be the characteristics of either a new or a disengaged player).

When we cross-check with the attributes in rows 7-10, we see that on average ClusterO players have been inactive for the most number of days (21 days of inactivity is the typical definition of churn, so these players are on the verge of churning). The highest revenue segment is Cluster 4 whose play characteristics interest- ingly, seem similartothe PCsegment in Section 6.1. Cluster 1 is the mostskillful segment, which wasalso theconclusion drawn from game play. Apart from these personas, players' playing behaviors may evolve over time. These changes in playing patterns can be pre-cursor to various important events pertaining to business. In this section, we show this with the exam pie of player churn that can be captured by longitudinal analysis of playing behavior (and its variations).

We define churn as players who have been inactive for more than 21 days. In order to motivate how the HQE model can help with churn prediction, we refer to Figure 11. (a) shows the drop action in the last 500 games of one player who churned due to long-term monetary disappointment; (b) shows the drop metric values in his last 100 games, moving averaged over 10 game windows. We see from 11(a) that in the last 100 games, the player has not dropped a single game, indicating onset of aggressiveness or disengagement. We also see from 11(b) that the average points has been steadily increasing, which is quite indicative of long-term monetary disappointment.

We have an existing XG Boost model to predict players who are about to churn. The features used in this model include number of active days, bet a mount, win a mount, revenue generated, deposit amount in the wallet, a mount withdrawn (the wins that the player can take out), and the change in these data-points over different time windows. We augment this model, by including the drop metrics and their changes over time as additional features. We observe that the AUC improves by 8.9%from 0.614to 0.669.

Figure 12 shows the drop action (a) and drop metrics (b) of one player who was not flagged by the baseline model, but was predicted with drop metrics included. We see from 12 (b) that about 100 games before he churned, there is a steep rise in the mistake ratio, drop proba bility and average points. This indicates a streak of bad hands, and bad decisions (there is also a decrease in the drop ratio in intermediate hands), resulting in churn due to disengagement.

As submitted above , as the game progresses, the hand of each player changes in each round, with the discard stack growing, and the closed stack decreasing (on an average, a game of rummy lasts for about 5 to 6 rounds). A strategic player will closely keep track of which cards are being discarded and picked by opponents and continually estimate the set of cards remaining in the closed deck in order to plan his melds accordingly.

So often, a GAIM model will have to read multiple images as input, and also keep track of past events.

To demonstrate this , using the hand representation and HICU as building block, the present invention proposes (i) Multi-input Win Likelihood Model (MWL), to predict the win probability of a player given a snapshot of the game at an intermediate round (i.e., takes multiple images as input - his hand, the hands of all his opponents and the current closed stack), and (ii) Sequential Win Likelihood Model (SWL), an LSTM- based model which predicts the win likelihood given the sequence of a player's hands (i.e., hand image from Round 1, Round 2 and Round 3).

The present section of the description showcases the versatility the GAI M framework in terms of information that can be mined. As with the HQE model, these models can be used for not just predicting the game outcome, but also to understand finer nuances of game evolution.

Further, the model (Multi-Input Win Likelihood Model (MWL)) takes n + 1 images as inputs, where n is the number of players in the game, along with the closed stack image (discard stack is omitted because it would simply be the compliment of close stack and all players' hands). The architecture of the model along with sample input images are shown in Figure 13 for two player games where each player's hand and the closed stack images are fed in parallel to separate pipelines of the HICU unit. The outputs of these are fed to two dense layers of 1024 and 512 nodes, ending with a softmax layer for binary classification. The ground truth label is 1 if the first (top-most) hand wins at the end of the game, and 0 otherwise (for each n-player game, there are n records in the training set, for each player's outcome, with the order of players preserved). Table 11 compares the performance with three other variations of the model : ) only the player's hand as input, (ii) the player's hand and closed stack as input, (iii) player and his opponent's hands as input, (iv) player and opponent's hands along with closed stack as input (all at the end of Round 3). We see that, as expected, the performance is best when most information provided.

As another illustration of the versatility of the framework, via Sequential Win Likelihood Model (SWL). , it is discloses how the evolution of a player's hand can be captured with a simple LSTM model.

The win likelihood of a player is modelled given his hands in Round 1, Round 2 and Round 3, i.e., we classify a sequence of hands as likely to win or not.

The architecture of the model (Figure 13(b)) consists of Time Distributed layers of timestamp=3 for each layer of HICU, the output of which is fed to LSTM with 64 nodes. The results are shown in Table 11 and it outperforms single HICU-based models that use only R1 hand and only R3 hand. As expected, there is only a slight improvement when the 3-step sequence is classified, compared to using only R3 hands, for win prediction.

However, this is evidence of how sequential information can be learned via GAIM. For instance, going beyond just win prediction, such a sequential model can allow to benchmark and identify at which point during the game evolution, a player has deviated from his usual (or best) game play behaviour.

Thus, present invention proposes GAIM, a framework that enables benchmarking and profiling of players with respect to what game actions they take in response to different game states.

The present invention shows that multi-player card games can be effectively represented in an image format, and custom CNN model can be used to estimate the hand quality (with the first drop probability as its proxy). The present invention also discloses that the model is a near-true state model, where misclassifications mostly occur due to either player mistakes or genuinely hard hands.

The present invention can be used for player behavioural analyses in response to various game states and evolution. Future work includes enriching sequential evolution of the game to identify deviations in player behaviour within a game and root cause analyses of player disengagement.

Such technical solution allows to create a generic framework to be applied for a detailed analysis on the image.

Additionally, proposed solution does not require any manual intervention in representing complex semantic features from the GUI elements state information.

The present invention provides significant advantages over existing feed forward neural networks, which are not able to capture comprehensive state information as such know approaches take inputs in a linear sequence of bits.

In contrast, the present invention encapsulates the state information of GUI elements. For Example, the encapsulated information includes a set of GUI tabs for interaction with the users along with additional meta-data information in the form of a multi-dimensional array in such a way that the semantics of the user behaviour of the user are learnt by the model of the proposed present invention.

Although the preferred embodiment of the present invention has been described, it will be apparent to those skilled in the art that many changes and modifications may be made without departing from the invention in its broader aspects.